Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Real business cycles The Real Business Cycle model has become the dominant mode of business analysis within the new classical school of macroeconomic thought. It has been the focus of a great deal of debate and controversy, and yet, to date, there has been no single source for material on real business cycles, much less one which provides a balanced account of both sides of the debate. This volume fills the gap by presenting a collection of the essential articles that define and develop the real business cycle school, as well as those that criticize it. Key areas covered include: • • • • • the establishment of the real business cycle program the aims and methods of the real business cycle school analysis of the statistics and econometrics of the calibration techniques advocated by real businesss cycle modelers assessment of the empirical success of the real business cycle model from a variety of methods and perspectives the measurement of technology shocks in business cycle models (the Solow residual). A detailed Introduction assesses the strengths and weaknesses of real business cycle theory from a critical perspective and includes a nontechnical User’s Guide to the formulation and solution of models which will aid understanding of the articles and issues discussed. Offering a thorough assessment of the real business cycle program, this volume will be a vital resource for students and professionals in macroeconomics. James E.Hartley is Assistant Professor of Economics at Mount Holyoke College, Massachusetts. Kevin D.Hoover is Professor of Economics at the University of California, Davis. Kevin D.Salyer is Associate Professor of Economics, also at the University of California, Davis. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors Real business cycles A Reader Edited by James E.Hartley, Kevin D.Hoover, and Kevin D.Salyer London and New York © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors First published 1998 by Routledge 11 New Fetter Lane, London EC4P 4EE This edition published in the Taylor & Francis e-Library, 2006. “To purchase your own copy of this or any of Taylor & Francis or Routledge’s collection of thousands of eBooks please go to www.eBookstore.tandf.co.uk.” Simultaneously published in the USA and Canada by Routledge 29 West 35th Street, New York, NY 10001 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors All rights reserved. No part of this book may be reprinted or reproduced or utilized in any form or by any electronic, mechanical, or other means, now known or hereafter invented, including photocopying and recording, or in any information storage or retrieval system, without permission in writing from the publishers. British Library Cataloguing in Publication Data A catalogue record for this book is available from the British Library Library of Congress Cataloging in Publication Data Real business cycles/[edited by] James E.Hartley, Kevin D.Hoover, and Kevin D.Salyer. p. cm. Includes bibliographical references and index. 1. Business cycles. I. Hartley, James E., 1966– . I I. Hoover, Kevin D., 1955– . III. Salyer, Kevin D., 1954– . H B3711.R35 1998 97–40397 CI P I S BN 0-203-07071-2 Master e-book I S B N ISBN 0-203-22306-3 (Adobe eReader Format) ISBN 0-415-16568-7 (hbk) IS BN 0-415-17154-7 (pbk) © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors “It is thus that the existence of a common standard of judgment leads physicists, who are no more saintly than economists, to question their own best work.” Steven Weinberg © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors Contents Acknowledgements xi Part I Introduction 1 The Limits of Business Cycle Research 2 A User’s Guide to Solving Real Business Cycle Models 3 43 Part II The foundations of real business cycle modeling 3 4 5 Finn E.Kydland and Edward C.Prescott, “Time to build and aggregate fluctuations,” Econometrica 50(6), November 1982, pp. 1345–1369. 57 Edward C.Prescott, “Theory ahead of business cycle measurement,” Federal Reserve Bank of Minneapolis Quarterly Review 10(4), Fall 1986, pp. 9–22. 83 Lawrence H.Summers, “Some skeptical observations on real business cycle theory,” Federal Reserve Bank of Minneapolis Quarterly Review 10(4), Fall 1986, pp. 23–27. 97 6 Edward C.Prescott, “Response to a skeptic,” Federal Reserve Bank of Minneapolis Quarterly Review 10(4), Fall 1986, pp. 28–33. 102 7 Robert G.King, Charles I.Plosser and Sergio T.Rebelo, “Production, growth, and business cycles I: The basic neoclassical model,” Journal of Monetary Economics 21(2), March 1988, pp. 195–232. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 108 viii CONTENTS Part III Some extensions 8 Gary D.Hansen, “Indivisible labor and the business cycle,” Journal of Monetary Economics 16(3), November 1985, pp. 309–328. 149 9 Gary D.Hansen and Randall Wright, “The labor market in real business cycle theory,” Federal Reserve Bank of Minneapolis Quarterly Review, Spring 1992, pp. 2–12. 168 10 Lawrence J.Christiano, and Martin Eichenbaum, “Current real business cycle theories and aggregate labor market fluctuations,” American Economic Review 82(3), June 1992, 430–450. 179 11 Thomas F.Cooley and Gary D.Hansen, “The inflation tax in a real business cycle model,” American Economic Review 79(4), September 1989, pp. 733–748. 200 Part IV The methodology of equilibrium business cycle models 12 Finn E.Kydland and Edward C.Prescott, “The econometrics of the general equilibrium approach to business cycles,” Scandinavian Journal of Economics 93(2), 1991, pp. 161–178. 219 13 Finn E.Kydland and Edward C.Prescott, “The computational experiment: An econometric tool,” Journal of Economic Perspectives 10(1), Winter 1996, pp. 69–86. 237 14 Lars Peter Hansen and James J.Heckman, “The empirical foundations of calibration,” Journal of Economic Perspectives 10(1), Winter 1996, pp. 87–104. 254 15 Kevin D.Hoover, “Facts and artifacts: Calibration and the empirical assessment of real-business-cycle models,” Oxford Economic Papers 47(1), March 1995, pp. 24–44. 272 Part V The critique of calibration methods 16 Allan W.Gregory and Gregor W.Smith, “Calibration as testing: Inference in simulated macroeconomic models,” Journal of Business and Economic Statistics, 9(3), July 1991, pp. 297–303. 295 17 Mark W.Watson, “Measures of fit for calibrated models,” Journal of Political Economy 101(6), December 1993, pp. 1011–1041. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 302 CONTENTS ix 18 Fabio Canova, “Statistical inference in calibrated models,” Journal of Applied Econometrics 9, 1994, pp. 123–144. 333 19 Fabio Canova, “Sensitivity analysis and model evaluation in simulated dynamic general equilibrium economies,” International Economic Review 36(2), May 1995, pp. 477–501. 355 Part VI Testing the real business cycle model 20 Finn E.Kydland and Edward C.Prescott. “Business cycles: Real facts and a monetary myth,” Federal Reserve Bank of Minneapolis Quarterly Review 14(2), Spring 1990, pp. 3–18. 383 21 Sumru Altuǧ, “Time-to-build and aggregate fluctuations: Some new evidence,” International Economic Review 30(4), November 1989, pp. 889–920. 399 22 Fabio Canova, M.Finn, and A.R.Pagan, “Evaluating a real business cycle model,” in C.Hargreaves (ed.) Nonstationary Time Series Analysis and Cointegration. Oxford: Oxford University Press, 1994, pp. 225–255. 431 23 Robert G.King and Charles I.Plosser. “Real business cycles and the test of the Adelmans,” Journal of Monetary Economics 33(2), April 1989, pp. 405–438. 462 24 James E.Hartley, Kevin D.Salyer and Steven M.Sheffrin, “Calibration and real business cycle models: An unorthodox experiment,” Journal of Macroeconomics 19(1), Winter 1997, pp. 1–17. 496 25 Martin Eichenbaum, “Real business-cycle theory: Wisdom or whimsy,” Journal of Economic Dynamics and Control 15(4), October 1991, 607–626. 513 26 Gary D.Hansen and Edward C.Prescott. “Did technology shocks cause the 1990–1991 recession?” American Economic Review 83(2), May 1993, pp. 280–286. 533 Part VII The Solow residual 27 Robert M.Solow, “Technical change and the aggregate production function,” Review of Economics and Statistics 39(3), August 1957, pp. 312–320. 543 28 N.Gregory Mankiw, “Real business cycles: A new Keynesian perspective,” Journal of Economic Perspectives 3(3), Summer 1989, pp. 79–90. 552 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors x CONTENTS 29 Zvi Griliches, “The discovery of the residual: A historical note,” Journal of Economic Literature 34(3), September 1996, pp. 1324–1330. 564 30 Timothy Cogley and James M.Nason, “Output dynamics in real business cycle models,” American Economic Review 85(3), June 1995, pp. 492–511. 571 Part VIII Filtering and detrending 31 Robert J.Hodrick and Edward C.Prescott, “Postwar US business cycles: An empirical investigation,” Journal of Money, Credit and Banking 29(1), February 1997, pp. 1–16. 593 32 A.C.Harvey and A.Jaeger. “Detrending, stylized facts and the business cycle,” Journal of Applied Econometrics 8(3), 1993, pp. 231–247. 609 33 Timothy Cogley and James M.Nason. “Effects of the Hodrick-Prescott filter on trend and difference stationary time series: Implications for business cycle research,” Journal of Economic Dynamics and Control 19(1–2), January–February 1995, pp. 253–278. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 626 Acknowledgements The editors have benefited from the help of a number of people in putting this volume together. We thank: John Muellbauer for comments on a draft of Chapter 1; Michael Campbell for bibliographic work; Jeannine Henderson for her careful work in preparing camera-ready copy of the articles reproduced here; Alan Jarvis, Alison Kirk, and Laura Large of Routledge for their encouragement and help with the production details; Colin Cameron, Timothy Cogley, Martin Eichenbaum, Robert Feenstra, Andrew Harvey, Dale Heien, Richard Howitt, Finn Kydland, Martine Quinzii, Steven Sheffrin, and Gregor Smith for helping us to obtain copies of the articles suitable for the reproduction in the volume. We also acknowledge the following permissions to reprint: The Econometric Society and Dickey and Fuller for a table from “Liklihood Ratio Statistics for Auto-regression Time Series with a Unit Root” in Econometrica, 49, 1981, pp. 12–26. The Econometric Society and N.E. Savin and K.J.White for a table from “The Durbin-Watson Test for Serial Correlation with Extreme Small Samples or Many Regressors”, Econometrica, 45, 1977, 1989–1986 as corrected by R.W.Farbrother, Econometrica, 48, September 1980, p. 1554. The American Economic Association and Finn E.Kydland and Edward C.Prescott for “The Computational Experiment: An Econometric Tool” in Journal of Economic Perspectives, vol. 10:1, Winter 1996, pp. 69–86. The American Economic Association and Lars Peter Hansen and James H.Heckman for “The Empirical Foundations of Calibration” in Journal of Economic Perspectives, vol. 10:1, Winter 1996, pp. 87–104. The American Economic Association and Gary D.Hansen and Edward C.Prescott for “Did Technology Shocks Cause the 1990–1991 Recession” in American Economic Review, vol. 83:2, May 1993, pp. 280–286. The American Economic Association and Lawrence J. Christiano and Martin Eichenbaum for “Current Real-Business Cycle Theories and Aggregate Labour Market Functions” in American Economic Review, vol. 82:3, June 1992, pp. 430–450. The American Economic Association and © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors XII ACKNOWLEDGEMENTS Thomas F.Cooley and Gary D.Hansen for “The Inflation Tax in a Real Business Cycle Model” in American Economic Review, vol. 79:4, September 1989, pp. 733–748. The Econometric Society and Finn E. Kydland and Edward C.Prescott for “Time to Build and Aggregate Fluctuations”, in Econometrica, vol. 50:6, November 1982, pp. 1345–1369. Reed Elsevier Plc and Edward C.Prescott for “Theory Ahead of Business Cycle Measurement” in Federal Reserve Bank of Minneapolis Quarterly Review, vol. 10:4, Fall 1986, pp. 9–22. Reed Elsevier Plc and Lawrence H. Summers for “Some Sceptical Observations On Real Business Cycle Theory” in Federal Reserve Bank of Minneapolis Quarterly Review, vol. 10:4, Fall 1986, pp. 23–27. Reed Elsevier Plc and Edward C.Prescott for “Response to a Sceptic” in Federal Reserve Bank of Minneapolis Quarterly Review, vol. 10:4, Fall 1986, pp. 28–33. Elsevier Science BV, The Netherlands and Robert G.King, Charles I.Plosser and Sergio T.Rebelo for “Production, Growth and Business Cycles I: The Basic Neoclassical Model” in Journal of Monetary Economics, vol. 21:2, March 1988, pp. 195–232. Elsevier Science BV, The Netherlands and Gary D.Hansen for “Indivisible Labour and the Business” in Journal of Monetary Economics, vol. 16:3, November 1985, pp. 309–328. Reed Elsevier Plc and Gary D. Hansen and Randall Wright for “The Labour Market in Real Business Cycle Theory” in Federal Reserve Bank of Minneapolis Quarterly Review, Spring 1992, pp. 2–12. Basil Blackwell Ltd and Finn E.Kydland and Edward C.Prescott for “The Econometrics of the General Equilibrium Approach to Business Cycles” in Scandinavian Journal of Economics, vol. 93:2, 1991, pp. 161–178. Oxford University Press Journals and Kevin D. Hoover for “Facts and Artifacts: Calibration and the Empirical Assessment of Real-Business Cycle Models” in Oxford Economic Papers, vol. 47:1, March 1995, pp. 24–44. The American Statistical Association and Allan W.Gregory and Gregor W.Smith for “Calibration as Testing: Inference in Simulated Macroeconomic Growth Models” in Journal of Business and Economic Statistics, vol. 9:3, July 1991, pp. 297–303. University of Chicago Press Journals and Mark W.Watson for “Measures of Fit for Calibrated Models” in Journal of Political Economy, vol. 101:6, December 1993, pp. 1011–1041. John Wiley & Sons Ltd and Fabio Canova for “Statistical Inference in Calibrated Models” in Journal of Applied Econometrics, 9, 1994, pp. 123–144. University of Pennsylvania and Fabio Canova for “Sensitivity Analysis and Model Evaluation in Simulated Dynamic General Equilibrium Economies” in International Economic Review, vol. 36:2, May 1995, pp. 477–501. Reed Elsevier Plc and Finn E.Kydland and Edward C.Prescott for “Business Cycles: Real Facts and a Monetary Myth” in Federal Reserve Bank of Minneapolis Quarterly Review, vol. 14:2, Spring 1990, pp. 3–18. University of Pennsylvania and Sumru Altu for “Time-to-Build and Aggregate Fluctuations”, in International Economic Review, vol. 30:4, November 1989, pp. 889–920. Fabio Canova, M.Finn, and A.R.Pagan for “Evaluating a Real Business Cycle Model” © Colin © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors ACKNOWLEDGEMENTS XIII P.Hargreaves 1994, reprinted from Nonstationary Time Series Analysis and Cointegration, edited by C.Hargreaves (1994) by permission of Oxford University Press. Elsevier Science BV, The Netherlands and Robert G.King and Charles I.Plosser for “Real Business Cycles and the Test of the Adelmans” in Journal of Monetary Economics, vol. 33:2, April 1989, pp. 405–438. Louisiana State University Press and James E.Hartley, Kevin D.Salyer and Steven M.Sheffrin for “Calibration and Real Business Cycle Models: An Unorthodox Experiment” in Journal of Macroeconomics, vol. 19:1, Winter 1997, pp. 1–17. Elsevier Science BV, The Netherlands and Martin Eichenbaum for “Real Business Cycle Theory: Wisdom or Whimsy”, Journal of Economic Dynamics and Control, vol. 15:4, October 1991, pp. 607–626. MIT Press Journals and Robert M.Solow for “Technical Change and the Aggregate Production Function” in Review of Economics and Statistics, 39, August 1957, pp. 312–320. The American Economic Association and N.Gregory Mankiw for “Real Business Cycles: A New Keynesian Perspective” in Journal of Economic Perspectives, vol. 3:3, Summer 1989, pp. 79–90. The American Economic Association and Zvi Griliches for “The Discovery of the Residual: A Historical Note”, Journal of Economic Literature, vol. 34:3, September 1996, pp. 1330–1334. The American Economic Association and Timothy Cogley and James M.Nason for “Output Dynamics in Real Business Cycle Models” in American Economic Review, vol. 85:3, June 1995, pp. 492–511. Ohio State University Press and Robert J.Hodrick and Edward C. Prescott “Postwar US Business Cycles: An Empirical Investigation”, in Journal of Money Credit and Banking, vol. 29:1, February 1997, pp. 1–16. John Wiley & Sons Ltd and A.C.Harvey and A.Jaeger for “Detrending, Stylized Facts and the Business Cycle” in Journal of Applied Econometrics, 8, 1993, pp. 231–247. Elsevier Science BV and Timothy Cogley and James M.Nason for “Effects of the HodrickPrescott Filter on Trend and Difference Stationary Time Series: Implications for Business Cycle Research”, Journal of Economic Dynamics and Control, vol. 19:1–2, January–February 1995, pp. 253–278. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors Part I Introduction © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors Chapter 1 The limits of business cycle research “That wine is not made in a day has long been recognized by economists.” With that declaration in Kydland and Prescott’s “Time to Build and Aggregate Fluctuations” (1982 [3]: 1345),* the real business cycle school was born. Like wine, a school of thought is not made in a day. Only after it has matured is it possible to judge whether it is good and to separate the palatable portions from the dregs. The literature on real business cycle models has now sufficiently aged, ready for the connoisseurs to pass judgment. To facilitate those judgments, we have collected together in this volume thirty-one previously published articles relevant to real business cycle models. While there has been no shortage of commentaries on the real business cycle program, the commentaries have been widely scattered and have often focused on narrow aspects of the models or represented partisan positions. Until now, there has not been an easily accessible means for students of business cycles to assess the real business cycle program on the basis of the original sources from the perspectives of the critics as well as the proponents. To date, the most systematic accounts of the real business cycle program are found in the works of active proponents, particularly in Thomas Cooley’s (ed.) Frontiers of Business Cycle Research (1995b), and in the programmatic manifestoes of Kydland and Prescott (1991 [12], 1996 [13]). Yet the critical literature is burgeoning. The present volume brings together the important articles which make the case for and against real business cycle models. The articles begin with the classics of the real business cycle school, starting with Kydland and Prescott’s (1982 [3]) seminal model. In addition, we include articles on the methodology of real business cycle models, particular aspects of the program (e.g., calibration, the measurement of technology shocks, * Throughout Chapters 1 and 2, the bold numbers in the square brackets within the text references refer to later chapters in this volume. However, all page numbers in these references are the page numbers from the original article. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 4 INTRODUCTION methods of detrending), as well as articles that attempt to evaluate the empirical success of the real business cycle model. The real business cycle program is still a very active one. We therefore hope that this anthology will prove useful to students and professional macroeconomists working on real business cycle models—bringing some perspective to the literature and pointing the way to further research. As an aid to research, the articles are reprinted here as facsimiles rather than reset. The preservation of the original pagination, facilitating authoritative citations, more than compensates, we believe, for the loss of an aesthetically pleasing typographical consistency. It is difficult for the neophyte in any area of economics to jump into the middle of a literature that was meant to advance the current interests of established economists, rather than a didactic purpose. In the remainder of this introductory chapter, we aim to provide a segue from the common knowledge of the advanced student of macroeconomics (or of the nonspecialist professional) to the essential elements of the real business cycle program. The objective is to provide a clear, accessible background to the literature that avoids unnecessary technical complications. At the same time, in this introductory essay we present our own assessment of the successes and failures of the real business cycle program. It is an assessment with which many economists will strongly disagree. We nevertheless hope that it will be easier for others to articulate their own assessments against the background of our effort. The articles reprinted in the volume provide the necessary raw materials. The technical demands of real business cycle models are often very high. As a further aid to the neophyte reader of the literature, and to the potential user of the models, the second introductory chapter to the volume is a user’s guide to real business cycle models, which provides a step-bystep account of how to formulate, solve, and simulate a real business cycle model. So much for preliminaries; let us turn now to the background of the real business cycle program and to the assessment of its successes and failures. I THE REAL BUSINESS CYCLE CONJECTURE The philosopher of science Karl Popper (1959, 1972) argued that science progresses through a series of bold conjectures subjected to severe tests. Most conjectures are false and will be refuted. The truth, by definition, will survive the ordeal of testing and emerge unrefuted at the end of inquiry in an infinitely distant future. The boldest conjectures are often the most fruitful, because, making the strongest claims, they are the most readily refuted and their refutation narrows the universe of acceptable conjectures most rapidly. We argue that real business cycle models are © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE LIMITS OF BUSINESS CYCLE RESEARCH 5 bold conjectures in the Popperian mold and that, on the preponderance of the evidence (to use a legal phrase), they are refuted. It is not, however, straightforward to see this, because the real business cycle conjecture is advanced jointly with a claim that models should be assessed using a novel strategy. We must therefore evaluate the conjecture and the assessment strategy simultaneously. Since the publication of Kydland and Prescott’s “Time to Build and Aggregate Fluctuations” (1982 [3]), the paradigm real business cycle model, a large and active group of new classical macroeconomists have elaborated and developed the real business cycle model. As important as these developments are to the real business cycle program, none of them fundamentally affects the critical points that we will make.1 Our assessment will, therefore, focus on the original Kydland and Prescott model and its successor models in a direct line. We will also refer frequently to the programmatic statements and methodological reflections of Kydland, Prescott and Lucas, the most articulate defenders of the aims and methods of equilibrium business cycle models. (i) Equilibrium business cycles To common sense, economic booms are good and slumps are bad. Economists have attempted to capture common sense in disequilibrium models: full employment is modeled as an equilibrium: that is, as a situation in which each worker’s and each producer’s preferences (given his or her constraints) are satisfied, while anything less than full employment represents a failure of workers or employers or both to satisfy their preferences. The real business cycle model is an extraordinarily bold conjecture in that it describes each stage of the business cycle—the trough as well as the peak—as an equilibrium (see, for example, Prescott, 1986a [4]: 21). This is not to say that workers and producers prefer slumps to booms. We all prefer good luck to bad.2 Rather it is to deny that business cycles represent failures of markets to work in the most desirable ways. Slumps represent an undesired, undesirable, and unavoidable shift in the constraints that people face; but, given those constraints, markets react efficiently and people succeed in achieving the best outcomes that circumstances permit. Some other models have come close to the real business cycle conjecture. Models of coordination failure treat booms and slumps as two equilibria and posit mechanisms that push the economy to one equilibrium or the other (e.g., Cooper and John, 1988; Bryant, 1983). Since the boom equilibrium is the more desirable, policies might seek to affect the mechanism in a way that improves the chances of ending in the boom state. The Phillips-curve models of Milton Friedman (1968) and Robert Lucas (1972, 1973) envisage people achieving their preferences conditional on an incorrect understanding of the true situation. Booms occur when © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 6 INTRODUCTION workers believe that real wages are higher than they really are, inducing them to supply more labor than they would if they knew the truth; slumps occur when workers believe that real wages are lower than they really are. Were people fully informed, there would be no booms or slumps.3 The real business cycle model is starker. As with Lucas’s monetary model, every stage of the business cycle is a Pareto-efficient equilibrium, but the source of the fluctuations is not misperceptions about prices or the money supply, but objective changes in aggregate productivity (so-called technology shocks). Thus, in the midst of a slump (i.e., a bad draw), given the objective situation and full information, every individual, and the economy as a whole, would choose to be in a slump. Contrary to the claims of some proponents of the real business cycle (e.g., Hodrick and Prescott, 1997 [31]: 1), there is no pre-Keynesian historical precedent for viewing business cycles as equilibria. Kydland and Prescott (1991 [12]) see such a precedent in the business cycle models of Ragnar Frisch (1933), while Lucas (1977:215; 1987:47 inter alia) sees such a precedent in the work of Hayek (1933, 1935) and other members of the Austrian School. Hoover (1988, ch. 10; 1995 [15]) demonstrates that these precedents are, at best, superficial. Frisch’s business cycle models are aggregative and do not involve individual optimization, even of a representative agent. Some Austrians reject the notion of equilibrium altogether. Hayek, who is not among these, accepts dynamic equilibrium as an ideal case, but sees business cycles as the result of mismatches of capital type and quantity to the needs of production transmitted to unemployment through a failure of wages and prices to adjust to clear markets in the short run—clearly a disequilibrium explanation.4 The real business cycle model advances a novel conjecture as well as a bold one. (ii) The novelty of the real business cycle model Novel in their bold conjecture, real business cycle models nonetheless have precursors. The primary antecedent is Robert Solow’s (1956, 1970) neoclassical growth model. In this model, aggregate output (Y) is produced according to a constant-returns-to-scale production function Φ(•) using aggregate capital (K), aggregate labor (L), and a production technology indexed by Z:5 (1.1) Consumption follows a simple Keynesian consumption function: (1.2) where s is the marginal propensity to save. Since Solow was interested in long-term growth, he ignored the aggregate demand pathologies that concerned earlier Keynesian economists and assumed that people’s plans © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE LIMITS OF BUSINESS CYCLE RESEARCH 7 were coordinated so that savings (S) equaled investment (I) ex ante as well as ex post: (1.3) Capital depreciates at rate δ and grows with investment: (1.4) where indicates the rate of change of capital. Labor grows exogenously at a rate n per cent per unit time, and labor-augmenting technology (Z) improves at a rate percent per unit time, so that effective labor grows at n+. Under these circumstances, the economy will converge to a steady state in which the growth of capital after compensating for depreciation is just enough to match the growth of effective labor. Along the steady-state growth path both capital and effective labor grow at a rate n+; and, since both inputs to production are growing at that steady rate, so is output itself. In the Solow growth model we need to distinguish between equilibrium and steady state. The model is always in equilibrium, because ex ante savings always equals ex ante investment (equation (1.3)). But the model need not be in steady state (i.e., growing at n+). Any change in the data that drives the economy away from the steady state (for example, a change in s or n) will also produce changes in capital and output (adjustments to a new steady state), but the economy remains in continuous equilibrium along the adjustment path. Lucas (1975) employed the Solow growth model to solve a difficulty in his own analysis of business cycles. Lucas (1972, 1973) explained the business cycle as the reaction of workers and producers to expectational errors induced by monetary policy. The difficulty was to explain why such expectational mistakes should not be corrected quickly so that business cycles were short week-to-week, month-to-month, or quarter-to-quarter fluctuations rather than the five- to six-year cycles typically observed. Lucas’s solution was to distinguish, in Ragnar Frisch’s (1933) useful terminology, between impulses that begin a business cycle and propagation mechanisms that perpetuate a cycle. Expectational errors were the impulses. These impulses drove the economy away from steady state. Ex post the economy was seen to be in disequilibrium until the expectational errors were corrected. But even when they had been corrected, the economy was returned to an equilibrium away from the steady state. The process of adjusting capital in order to regain the steady state would be a relatively slow one. This was the propagation mechanism. In keeping with the new classical agenda of reducing macroeconomics to microeconomic foundations, Lucas replaced the stripped-down demand © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 8 INTRODUCTION behavior of the Solow growth model with the assumption that the behavior of the aggregate economy can be described by the utility-maximizing choices of a representative agent, who chooses consumption and labor supply by solving a dynamic, intertemporal optimization problem. In effect, the simple consumption function (equation (1.2)) was replaced by a permanentincome (or life-cycle) consumption function, and investment was replaced by a neoclassical investment function in which the opportunity cost of capital determines the rate of investment. Unlike in the Solow model, the factors important to the savings decision now enter separately from those important to the investment decision. Aggregate demand pathologies are, nonetheless, impossible, because in Lucas’s model the same agents make both the savings and the investment decision, which insures ex ante coordination, and the agents have rational expectations, which insures that mistakes about the future course of the economy are necessarily unsystematic. Furthermore, the supply of labor responds elastically to temporarily high real wages: workers make hay while the sun shines. Kydland and Prescott’s (1982 [3]) seminal real business cycle model is a direct outgrowth of Lucas’s monetary growth model. It differs from the Lucas model in that there is no monetary sector; technology shocks (i.e., deviations of Z in equation (1.1) from trend) supply the impulse to business cycles. The model does not rely on expectational errors. There is no need. Lucas originally posited expectational errors as a way of permitting changes in the stock of money to have real effects on the economy without violating the assumption that money is neutral in the long run. In Kydland and Prescott’s model, technological change has real effects regardless of whether it is anticipated. While some of the variability in aggregate output, consumption, investment, and labor supply in Kydland and Prescott’s model is attributable to the unexpectedness of technology shocks, the aggregate variables would fluctuate even if technological change were perfectly anticipated. In a recent summary of the real business cycle methodology, Kydland and Prescott (1997:210) state that “we derive the business cycle implications of growth theory.” Seen in context, this is misleading. Historically, it is not the use of the growth model that distinguishes the real business cycle model from earlier business cycle models. Rather it is finding the impulses in technology shocks, and modeling the economy in continuous equilibrium. Both in their theoretical development and (as we shall see presently) in their quantitative implementation, real business cycle models abstract from the traditional concerns of growth theory. They provide no analysis of the steady-state rate of growth at all, but take the factors that determine it as exogenously given.6 Instead, the focus is on the deviations from the steady state. Only if growth theory were synonymous with aggregate general equilibrium models with an optimizing representative agent would it be fair to say that their behavior is the implication of © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE LIMITS OF BUSINESS CYCLE RESEARCH 9 growth theory. But Solow’s growth model is an example of an aggregate general equilibrium model that does not posit an optimizing representative agent. Technology shocks would be propagated in a Solow growth model, though rather slowly, for the convergence time to steady state is long in a realistically parameterized Solow model (cf. Sato, 1966). The characteristic business cycle behavior in real business cycle models comes from the shocks and from the optimizing model itself (of which more presently), rather than from the fact that these are embedded in a growth model. (iii) A quantified idealization Real business cycle models are implemented by giving specific functional forms to the equations of the optimal growth model. This is most easily seen for production; equation (1.1) is replaced by a specific function, very often the Cobb-Douglas production function: (1.1´) where is the share of labor in national output. An equation such as (1.1´) could be estimated as it stands or jointly with the other equations in the model to determine the value of . Real business cycle proponents do not typically estimate the parameters of their models. Instead, they assign values to them on the basis of information from sources outside the model itself. This is known as calibration of the model. The value chosen for is usually the average value that the labor share takes in suitably adapted national accounts.7 The value of the depreciation rate (␦) is calibrated similarly. As we mentioned already equations (1.2) and (1.3), which represent aggregate demand in the Solow growth model, are replaced in real business cycle models by an optimization problem for a representative agent who is both consumer and producer. The representative agent maximizes a utility function: (1.5) subject to current and future production constraints given by equation (1.1´), and linked together by equation (1.4), which governs the evolution of the capital stock. The set {Ct} is the set of current and future levels of consumption, and {Lt} is the set of current and future supplies of labor (the time subscript t=0, 1, 2, …, ∞). The utility function must be calibrated as well. This is usually done with reference to the parameters estimated in unrelated microeconomic studies.8 The calibrated model is nonlinear. To solve the model its equations are typically reformulated as linear approximations around the unknown steady state. This is the technical sense in which real business cycle models abstract from the concerns of traditional growth theory; for no explanation © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 10 INTRODUCTION of the steady state is sought—the focus is on (equilibrium) deviations from the steady state. The solution to the linearized model is a set of linear equations for output, consumption, labor, and investment of the form: (1.6–1) (1.6–2) (1.6–3) (1.6–4) where the lower-case letters are the deviations from steady state of the logarithms of the analogous upper-case variables. The coefficients γij are combinations of the calibrated parameters determined by solving the model.9 The right-hand variables in equations (1.6–1) to (1.6–4) are called state variables. They summarize the past evolution of the model economy and are exogenous in the sense that the representative agent takes them as given data and conditions his choices upon them (z is exogenous and k is determined from choices made in previous periods). Equations (1.6–1) to (1.6–4) detail the outcomes of those choices, that is, how the preferences of the representative agent interact with the constraints he faces, including the current state of z and k, to determine output, capital, labor, and investment. In the original Kydland and Prescott (1982 [3]) model, the technology shock, z, was modeled as a random process with parameters chosen to cause the model to mimic the variance of GNP in the US economy. Since z was artificial, there was no chance of direct period-by-period or historical comparisons of the modeled time series in equations (1.6–1) to (1.6–4) with their real-world equivalents. Kydland and Prescott, however, wished only to compare the covariances among the modeled time series to those among the actual series, so this did not seem problematic. Nevertheless, as Lucas (1987:43–45; cf. Prescott, 1986b [6]: 31) noticed, constructing the predicted output series to mimic actual output does not provide an independent test of the model.10 Beginning with Prescott (1986a [4]), real business cycle models have taken a different tack (cf. Kydland and Prescott, 1988). Solow (1957 [27]) attempted to quantify technical change by using a production function with constant returns to scale (such as equation (1.6)) to compute Z. Typically, real business cycle models use the CobbDouglas production function (equation (1.1´)) as follows: (1.7) This empirical measure of the technology parameter is known as the Solow residual. When estimated using actual data, the Solow residual, like the series used to compute it, has a trend (implying ⫽0), and so must be detrended before being used as an input to the real business cycle model. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE LIMITS OF BUSINESS CYCLE RESEARCH 11 Detrended log(Z) is the state-variable z.11 (iv) The limits of idealization The real business cycle model does not present a descriptively realistic account of the economic process, but a highly stylized or idealized account. This is a common feature of many economic models, but real business cycle practitioners are bold in their conjecture that such models nevertheless provide useful quantifications of the actual economy. While idealizations are inevitable in modeling exercises, they do limit the scope of the virtues one can claim for a model. In particular, the real business cycle program is part of the larger new classical macroeconomic research program. Proponents of these models often promote them as models that provide satisfactory microfoundations for macroeconomics in a way that Keynesian models conspicuously fail to do (e.g., Lucas and Sargent, 1979). The claim for providing microfoundations is largely based on the fact that new classical models in general, and real business cycle models in particular, model the representative agent as solving a single dynamic optimization problem on behalf of all the consumers, workers, and firms in the economy. However, the claim that representative agent models are innately superior to other sorts of models is unfounded. There is no a priori reason to accord real business cycle models a presumption of accuracy because they look like they are based on microeconomics. Rather, there are several reasons to be theoretically skeptical of such models.12 Most familiar to economists is the problem of the fallacy of composition, which Samuelson’s (1948) introductory economics text prominently addresses. It is difficult to deny that what is true for an individual may not be true for a group, yet, representative agent models explicitly embody the fallacy of composition. The central conceptual achievement of political economy was to analogize from the concerns of Robinson Crusoe—alone in the world—to those of groups of people meeting each other in markets. The complexities of economics from Adam Smith’s invisible hand to Arrow and Debreu’s general equilibrium model and beyond have largely been generated from the difficulties of coordinating the behavior of millions of individuals. Some economists have found the source of business cycles precisely in such coordination problems. By completely eliminating even the possibility of problems relating to coordination, representative agent models are inherently incapable of modeling such complexities. Problems of aggregation are similar to problems arising from the fallacy of composition. Real business cycle models appear to deal with disaggregated agents, but, in reality, they are aggregate models in exactly the same way as the Keynesian models upon which they are meant to improve. The conditions under which a representative agent could © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 12 INTRODUCTION legitimately represent the aggregate consequences of, and be deductively linked to, the behavior individuals are too stringent to be fulfilled: essentially all agents must be alike in their marginal responses.13 Because it is impracticable, no one has ever tried to derive the aggregate implications of 260 million people attempting to solve private optimization problems. The real business cycle model thus employs the formal mathematics of microeconomics, but applies it in a theoretically inappropriate circumstance: it provides the simulacrum of microfoundations, not the genuine article. It is analogous to modeling the behavior of a gas by a careful analysis of a single molecule in vacuo, or, of a crowd of people by an analysis of the actions of a single android. For some issues, such models may work well; for many others, they will miss the point completely.14 A significant part of the rhetorical argument for using real business cycle methodology is an appeal to general equilibrium theory. However, because the models do not reach a microfoundational goal of a separate objective function for every individual and firm, the models are at best highly idealized general equilibrium models. Real business cycle theorists do not appear to be aware of the degree to which this undermines certain sorts of claims that can be made for their models. The fact that they do not provide genuine microfoundations essentially removes any prior claims that real business cycle models are superior to Keynesian or other aggregate models. It is not difficult to understand why general equilibrium theory has such allure for economists in general and macroeconomists in particular. The theory provides for an extensive model of the economy with individual consumers maximizing utility and individual firms maximizing profits, all interacting in competitive markets, and despite all this complexity, it can be shown that an equilbrium exists. However, knowing that an equilibrium point exists is all well and fine, but it doesn’t get you very far. What else can we tell about the economy from the general equilibrium framework? The answer to that question turned out to be quite depressing; as Kirman (1989) subtitled a paper about this state of affairs, “The Emperor Has No Clothes.” After showing that an equilibrium point existed, people became interested in the question of whether it could be shown that the equilibrium was either unique or stable. In order to answer this question, the shape of the aggregate excess demand curve had to be determined. In a remarkable series of papers, Sonnenschein (1973, 1974), Mantel (1974, 1976), Debreu (1974) and Mas-Colell (1977) showed that in an economy in which every individual has a well-behaved excess demand function, the only restrictions on the aggregate excess demand function are that it is continuous, homogenous of degree zero in prices, and satisfies Walras’ Law. Nothing else can be inferred. Any completely arbitrary function satisfying those three properties can be an aggregate excess demand function for an economy of © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE LIMITS OF BUSINESS CYCLE RESEARCH 13 well-behaved individuals. Having an economy in which every single agent obeys standard microeconomic rules of behavior tells us virtually nothing about the aggregate economy. For example, not even something as basic as the Weak Axiom of Revealed Preference carries over from the microeconomic level to the macroeconomic level. (See Shafer and Sonnenschein (1982) for a complete, technical discussion of this literature and Kirman (1989) or Ingrao and Israel (1990) for good nontechnical discussions.) The implication can be stated in two ways. Even if we know that the microeconomy is well behaved, we know very little about the aggregate excess demand function. Or, given a completely arbitrary aggregate excess demand function satisfying the three characteristics above, we can find a well-behaved microeconomy that generates that aggregate function. Kirman’s (1992) article in the Journal of Economic Perspectives was largely centered on showing how these results invalidated the use of a representative agent model. There is simply no theoretical justification for assuming that the excess demand function of a representative agent bears any resemblance to the excess demand function for an aggregate economy. If we want to justify the notion that macroeconomics needs microfoundations by pointing to general equilibrium theory, then these results derived by general equilibrium theorists unambiguously demonstrate that the representative agent is flawed. Oddly, we seem to be simultaneously seeing a situation in which macroeconomists point to general equilibrium theory as a justification for representative agent models at the same time as general equilibrium theorists are prominently noting that the representative agent has no home in the theory. Thus the implicit claim in real business cycle theory that their representative agent models provide rigorous microfoundations is incorrect. Starting with first principles, or general equilibrium theory, only, we can derive all sorts of macroeconomics. Some form of aggregate structure must be provided. Beyond this, Kydland and Prescott argue that the models are designed to capture some features of the economy while ignoring or even distorting other features. They hold this to be one of their virtues, and argue that their failure to capture features that they were not designed to model should not count against them (Kydland and Prescott, 1991 [12]). We take this claim seriously. It should, nevertheless, be noted that it undermines the argument that we trust the answers which the models give us on some dimensions because they have been successful on other dimensions (Lucas, 1980:272). Kydland and Prescott (1996 [13]: 72) make exactly this claim with regard to using the Solow growth model to explore the business cycle. However, if the dimensions on which we need answers are ones on which, because of their idealized natures, the models are false, the success on other dimensions is irrelevant. As a point of logic, rigorous deductions are useful only if they start with true premises.15 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 14 INTRODUCTION Idealized models are useful because they are tractable, but only if they remain true in the features relevant to the problem at hand. Kydland and Prescott want idealized real business cycle models to provide quantitative conclusions about the economy. There is nothing in their construction that insures that they will succeed in doing so. Thus, part of the boldness of the real business cycle conjecture is the seriousness with which it takes the idealization of a representative agent. Although economists, at least since Alfred Marshall, have sometimes used representative agents as a modeling tool, new classical (and real business cycle) models expect the representative agent to deliver far more than earlier economists thought possible. For example, Friedman’s (1957) explication of the permanent-income hypothesis begins with something that looks like a representative agent, but Friedman uses the agent only as a means of thinking through what sorts of variables belong in the aggregate consumption function. He makes no attempt to derive an aggregate consumption function from his agent; in fact, he takes pains to note how different the aggregate function will look from the individual’s function. Real business cycle models, on the other hand, take the functions of the representative agent far more seriously, arguing that “we deduce the quantitative implications of theory for business cycle fluctuations” (Kydland and Prescott, 1997:211). However, for the reasons described above, these deductions are not the rigorous working out of microeconomic principles combined with a serious analysis of heterogeneity and aggregation. There is nothing in the construction of real business cycle models which insures that they will succeed in providing accurate quantitative conclusions. There is nothing that guarantees a priori their superiority. The proof of the pudding is in the eating: the real business cycle model must be tested and evaluated empirically. II TESTING (i) What are the facts about business cycles? Before real business cycle models can be tested, we must know precisely what they are meant to explain. Following Prescott (1986a [4]), advocates of real business cycle models have redefined the explanandum of business cycles. As observed already, common sense and the traditional usage of most economists holds that recessions are periods in which the economy is suboptimally below its potential. Business cycle theory has thus traditionally tried to explain what causes output to fall and then rise again. To be sure, this is not just a matter of output declining: when output declines, one expects employment, income, and trade to decline as well, and these declines to be widespread across different sectors.16 Nevertheless, the central fact to be explained was believed to be the decline and the subsequent recovery, and not the comovements of aggregate time series. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE LIMITS OF BUSINESS CYCLE RESEARCH 15 Even before the first real business cycle models, new classical macroeconomics shifted the focus to the comovements. Sargent (1979:256) offers one definition: “the business cycle is the phenomenon of a number of important economic aggregates (such as GNP, unemployment and layoffs) being characterized by high pairwise covariances at the low business cycle frequencies [two- to eight-year range]…. This definition captures the notion of the business cycle as being a condition symptomizing the common movements of a set of aggregates.” Lucas (1977:217) argues that the movements of any single economic aggregate are irregular, and “[t]hose regularities which are observed are in the comovements among different aggregative time series.”17 Real business cycle models view the business cycle in precisely the same way as Sargent and Lucas. The things to be explained are the correlations between time series, and the typical assessment of the success or failure of a model is to compare the correlations of the actual time series to those that result from simulating the model using artificially generated series for the technology shock (Z). Formal statistical measures of the closeness of the model data to the actual data are eschewed. Prescott (1986a [4]), for example, takes the fact that the model approximates much of the behavior of the actual aggregates as an indicator of its success. In the case in which the model data predict an empirical elasticity of output to labor greater than the theory, Prescott (1986a [4]: 21) argues “[a]n important part of this deviation could very well disappear if the economic variables were measured more in conformity with theory. That is why I argue that theory is now ahead of business cycle measurement.” Kydland and Prescott (1990 [20]) make similar arguments in opposing “business cycle facts” to “monetary myths.” For example, the real business cycle model predicts that the real wage is procyclical: a positive technology shock (an upward shift of the production function), which is a positive impulse to output, increases the marginal product of labor, and workers are paid their marginal products. In contrast, monetary business cycle models (Keynesian and monetarist) predict countercyclical real wages: a monetary shock, which is a positive impulse to output, increases aggregate demand and therefore the demand for labor, which requires a lower marginal product of labor (a movement along the production function). Kydland and Prescott (1990 [20]: 13–14) argue that a correlation of 0.35 between money lagged one period and current output is too low to support the view that money leads output; while a correlation of 0.35 between the real wage and output is high enough to support the view that the real wage is procyclical. They argue that if measurements were made in closer conformity to theory, the second correlation would be higher.18 But, even as it stands, they take the “business cycle facts” as supporting the real business cycle model. Theory may be ahead of measurement. It is well understood that as © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 16 INTRODUCTION science progresses, theoretical advances improve measurements. Thermometers were originally atheoretical devices that rested on some simple untested assumptions, such as the linearity of the relationship between temperature and the expansion of materials. In time, as the theory of heat developed, better thermometers were possible because theoretical understanding permitted corrections for departures from the initial assumptions and new methods of measurement that, among other things, permitted the range of temperature measurements to be increased to near absolute zero, at one end, and to millions of degrees, at the other. Given the best data in the world, however, simply mimicking the data is a weak test. Indeed, logically it is fallacious to argue that theory A implies that data behave as B, data in fact behave as B, therefore A is true. This is the fallacy of affirming the consequent. It is a fallacy because there is nothing to rule out incompatible theories C, D, and E also implying B. Popper’s concentration on refutations is a reaction to this fallacy in the form in which it was exposed by David Hume (1739): there is no logic that allows one to move inductively from particular instances to a general rule; there is no inductive logic analogous to deductive logic. It is a correct inference that theory A implies data behavior B, data fail to behave as B, therefore A is false. At best, the data limit the class of theories that are acceptable. One learns very little from knowing that a theory mimics the data—especially if it was designed to mimic the data. One needs to know that the data cannot be mimicked by rival theories. Although real business cycle models are often shown (without any formal metric) to mimic actual data, they have rarely been tested against rivals.19 It is usually regarded as a more stringent test of a model that it performs well on a set of data different from the one used in its formulation. Most often this means that models are formulated on one sample and then tested against a completely different sample. Kydland and Prescott (1997:210) offer a different argument: real business cycle models are formulated using the “stylized facts” of long-run growth theory and are then tested, not against a completely different data set, but for their ability to mimic the short-run business cycle behavior of the same data. While there is clearly merit in deriving empirically supported implications of one set of facts for another, this particular test provides very weak support for the real business cycle model. Many models that are fundamentally different from the real business cycle model, in that they posit neither continuous equilibrium nor impulses arising from technology shocks, are consistent with the “stylized facts” of growth (e.g., the constancy of the labor share in national income or the constancy of the capital-output ratio).20 (ii) Do real business cycle models fit the facts? Although it is a weak test to check whether models mimic the facts, it is a © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE LIMITS OF BUSINESS CYCLE RESEARCH 17 useful starting point. The fact that real business cycle models are idealized presents some difficulties in judging them even on such a standard. As Kydland and Prescott (1991 [12]: 169) stress, the real business cycle model is unrealistic in the sense that it aims only to capture certain features of the data rather than to provide a complete explanation. The econometric ideal is to provide predictions of dependent variables on the basis of all of the relevant independent variables, so that whatever errors are left are truly random and independent of any omitted variables. In contrast, real business cycle models are driven by a single variable, the technology shock, and aim to explain the relationships among a number of series (as in equations (1.6–1) to (1.6–4)) on the basis of this single shock. The success of the model is to be judged, in Kydland and Prescott’s view, on its ability to capture selected correlations in the actual data. There is no claim that it will do well in explaining correlations it was not designed to capture; nor is there any claim that its errors will be truly random, either in being mean zero and symmetrically (e.g., normally) distributed or in being independent from omitted variables. The dilemma is this: Theories are interpretable, but too simple to match all features of the data; rich econometric specifications are able to fit the data, but cannot be interpreted easily. The coefficients of a statistically well-specified econometric equation indicate the effects on the dependent variable ceteris paribus of a change in the independent variables. In general, these effects depend in a complicated way on the parameters of the deep relations that connect the variables together and generate the observed data. Lucas (1976) in his famous “critique” of policy analysis noticed the lack of autonomy of econometrically estimated coefficients and argues, in particular, that the values of the coefficients would not remain stable in the face of changes in monetary and fiscal policy regimes. One solution to the Lucas critique might be to identify the complex structure of the estimated coefficients. Hansen and Sargent (1980) map out a strategy for doing this. Essentially, the model is taken to be true and used, in the classic manner of textbook econometric identification, to disentangle the “deep” parameters (i.e., the parameters of the theory) from the estimated coefficients.21 The central difficulty with this strategy as a means of providing support for real business cycle models is that it does not work. In the case in which the model imposes more relationships among the parameters than there are parameters to identify, the model is said to be overidentified. Statistical tests can be used to assess whether these redundant relationships can be rejected empirically. Altu (1989 [21]) estimated an econometric version of the real business cycle model and tested its overidentifying restrictions. They were clearly rejected. This should not be surprising. An idealized model abstracts from too many of the features of the world for the resulting specification to meet the econometric ideal. Not only is it likely that the errors will not show © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 18 INTRODUCTION irreducible randomness and the appropriate symmetry, but they are unlikely to be independent of omitted variables. One may, of course, add additional variables into the regression equations. An econometric specification with many free parameters (i.e., many independent variables) will invariably fit better than a calibrated business cycle model. But then one loses the mapping of the theory onto the estimated coefficients that helped to disentangle the deep parameters. Kydland and Prescott advocate a second solution: eschew econometric estimation altogether. They believe that the advantage of the calibrated model is that it refers to theoretically interpretable parameters, so that counterfactual experiments can be given precise meanings: for example, the effects of a change in the persistence of the technology shock or in the relative risk aversion of consumers or, in richer real business cycle models, of government-policy rules (e.g., tax rates) have precise analogues in the calibrated model. A good model, in Kydland and Prescott’s view, is unrealistic, in the sense that it will not fit the data in the manner of a statistically well-specified econometric model, but it will fit with respect to certain features of interest. Calibration and model structure are adjusted until the models do well against those features of the data that are of interest. The development of the labor market in early real business cycle models provides an illustration of the strategy. Table 1.1 reproduces from Hansen (1985 [8]) some statistics for actual data and data generated from simulating two real business cycle models. Model I is a simple model similar to Kydland and Prescott (1982 [3]) in which labor is supplied in continuously variable amounts. The standard deviations of hours worked and productivity are nearly equal in Model I; while, in the actual data, hours worked are over 50 percent more variable. Model II is a modification of Model I in which, to capture the fact that workers typically must either work a full day or not work, labor must be supplied in indivisible eighthour units. Model II was created in part as an attempt to add realism to capture a feature that was not well described in Model I. In fact, it succeeds rather too well: hours are nearly three times as variable as productivity in Model II. Further developments of the real business cycle model (see, e.g., Hansen and Wright, 1992 [9]) aim in part to refine the ability to mimic the data on this point. A serious case can be made for choosing Kydland and Prescott’s strategy for dealing with the Lucas critique and favoring idealized models at the expense of achieving the econometric ideal of complete description of the data (see Hoover, 1995 [15]). The gain is that one preserves theoretical interpretability—though only at the cost of a limited understanding of the actual economy. Real business cycle modelers might respond that the choice is between limited understanding and no genuine understanding at all. But this would be too glib. There are at least two barriers to declaring © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors Table 1.1. Summary statistics for actual US data and for two real business cycle models Source: Hansen (1985 [8]), table 1; standard deviations in parentheses. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 20 INTRODUCTION the triumph of the real business cycle approach on the basis of the methodological virtues of idealization. First, most of the assessments of the success or failure of real business cycle models have been made in the casual manner exemplified by our previous discussion of Hansen’s (1985 [8]) divisible and indivisible labor models, using data no more precise than that of Table 1.1. The standard is what might be called “aesthetic R2”: whether Models I or II in Table 1.1 are too far from the actual data or close enough is a purely subjective judgment without a good metric. One response might be that no formal metric is possible, but that a more rigorous subjective evaluation would go some way to providing the missing standards. King and Plosser (1989 [23]) take this tack. They revive the method of Adelman and Adelman (1959), first used to evaluate the Klein-Goldberger econometric macromodel. King and Plosser simulate data from a real business cycle model and evaluate it using the business cycle dating procedures developed by Burns and Mitchell at the National Bureau of Economic Research. These techniques aim to characterize the repetitive features of the economy by averaging over historical business cycles normalized to a notional cycle length.22 Both the actual data and the simulated data from the real business cycle model are processed using Burns and Mitchell’s procedures. King and Plosser observe that it is difficult to discriminate between these two sets of data. But they note that the results “leave us uncomfortable,” because the same claims can be made on behalf of the Keynesian Klein-Goldberger model. Despite the greater detail in this study compared to typical assessments of real business cycle models, it is still wedded to aesthetic R2. In a similar vein, Hartley, Salyer, and Sheffrin (1997 [24]) examine the ability of the standard informal methods of assessment of real business cycle models to discriminate between alternative accounts of the actual economy. Hartley et al. use the Fair macroeconometric model of the US economy, a model in the tradition of Keynesian macroeconomic forecasting models such as the Brookings model or the Federal Reserve-University of Pennsylvania-Social Science Research Council model, to simulate data for a “Keynesian” economy in which demand shocks and disequilibrium are important.23 Calibrating a real business cycle to be consistent with the relevant parameters of the Fair model, they ask whether the real business cycle model, which is driven by technology shocks (these are calculated from the simulated data from the Fair model) and continuous equilibrium, can mimic a “Keynesian” economy. They find out that it can to at least as high a degree as it mimics the actual economy on the usual standards used by real business cycle modelers. One interpretation of this result is that it is very bad news for the real business cycle model, because it shows that it has no power of discrimination; its key assumptions do not restrict the sort of economies it can fit. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE LIMITS OF BUSINESS CYCLE RESEARCH 21 A real business cycle modeler, however, might riposte that the Fair model is a typical Keynesian model with many free parameters, so that it gives a good statistical description of the economy, even as it fails to model the true underlying mechanisms. Thus the fact that the real business cycle model “works” for simulations from the Fair model means nothing more than that it works for the actual economy. To check this interpretation, Hartley et al. alter two key parameters, those governing the interest elasticities of money demand and investment, changes which makes the simulations of the Fair model (particularly, the cross-correlations stressed by real business cycle analysts) behave more like European economies than like the US economy (see Backus and Kehoe, 1992). The real business cycle model is poor at mimicking the data from the altered Fair model. One might conclude that the real business cycle model is, in fact, discriminating. However, for a modeling strategy that takes pride in its grounding in fundamental and universal economic theory (the Solow growth model is not country-specific), this is hardly an attractive conclusion. Although European business cycles may be substantially different from American business cycles because of important institutional differences, real business cycle models typically seek to explain business cycles abstracting from those very institutional details. A second barrier to declaring the triumph of the real business cycle model on the basis of the methodological virtues of idealization is that, even if idealized models cannot be expected to fit as well as traditional econometric specifications under the best of circumstances, the conclusion that econometric estimation is irrelevant to the real business cycle model would be unwarranted. Calibration might be regarded as a form of estimation (Gregory and Smith, 1990; 1991 [16]). The problem is how to judge the performance of calibrated models against an empirical standard. Watson (1993 [17]) develops an asymmetrical measure of goodness of fit that is useful for real business cycle models precisely because their idealized nature makes it likely that the errors in fitting them to actual data are systematic rather than random. Even using his goodness of fit measure, Watson fails to produce evidence of high explanatory power for real business cycle models. Kydland and Prescott’s (1991 [12]) objection to traditional econometrics, what they call the “systems of equations” approach, is that an idealized model will not provide the necessary restrictions to permit the accurate estimation of its own parameters on actual data, because of the many features of the data that it systematically and intentionally ignores. Canova, Finn and Pagan (1994 [22]) undertake a somewhat less demanding test. Where Altug (1989 [21]) had tested restrictions that were strong enough to identify all the parameters of the real business cycle model and, therefore, to eliminate the need for calibration, Canova et al. examine the implications of a previously calibrated real business cycle model for the dynamic © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 22 INTRODUCTION behavior of various time series. They observe that the various time series can be described by a vector autoregression (VAR) in which each series is regressed on its own lagged values and the lagged values of each of the other series. What is more, if the real business cycle model is an accurate description of the actual data, then a number of restrictions must hold among the estimated parameters of the VAR. The real business cycle model implies three sets of restrictions on the VAR of two distinct types. First, various time series should be cointegrated. Two series are cointegrated when a particular linear combination of them is stationary (i.e., when its mean, variance, and higher moments are constant), even though the series are not separately stationary. There are two sets of such cointegration restrictions: (1) the state variables (the analogues of Z and K, the non-detrended counterparts to the state variables in equations (1.6–1) to (1.6–4)) must stand in particular linear relationships; (2) state variables and predicted values for various time series (e.g., the left-hand variables in equations (1.6–1) to (1.6–4)) to must also stand in particular linear relationships. Finally, once one has accounted for the cointegrating relationships among these time series and concentrates on their behavior about their common trends, there is a third set of restrictions (second type), which are the particular implications of the real business cycle model for the parameters of the VAR. Canova et al. use a calibrated real business cycle model with a considerably richer specification than Kydland and Prescott’s early models to derive the necessary restrictions on the VAR. These restrictions are then compared to the data. Canova et al. show that the restrictions do not hold. A particularly interesting finding is that the real business cycle model imputes too much importance to the productivity shock. Canova et al.’s imposition of a specific numerical calibration of the real business cycle model might limit the generality of their results: it might be said that the real business cycle model is correct in principle, but Canova et al. have failed to calibrate it correctly. In defense of their test, their choice of parameters is not at all atypical. What is more, they examine a limited range of alternative choices of parameters, asking the question: What parameters would it take to make the model agree with the data? Their results along these lines, however, are not nearly as comprehensive as they would need to be to close the case. Eichenbaum (1991 [25]) examines the issue of parameter choice more systematically. He begins by noting that the numerical values of the underlying parameters used to calibrate a real business cycle model are simply estimates of the true values. We do not know the true values of things like the depreciation rate or the variance of the shock to the Solow residual. Instead, we estimate these numbers from sample data, and there is a sampling error associated with every estimate. (Hansen and Heckman, 1996 [14]: 95, make a similar argument.) © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE LIMITS OF BUSINESS CYCLE RESEARCH 23 Eichenbaum finds that altering most of the parameters within the range of their sampling error does little to alter the behavior of the real business cycle model. The notable exceptions are the parameters associated with the Solow residual, which have large standard errors. He finds that at standard levels of statistical significance (5 percent critical values), technology shocks may account for as little as 5 percent and as much as 200 percent of the variance in output. Eichenbaum’s results suggest that, even if real business cycle models had no other problems, we cannot reject the view that technology shocks in conjunction with a real business cycle model explain only a small fraction of aggregate fluctuations. Although not decisive, conventional econometric tests of real business cycle models are not kind to the theory. Canova et al.’s investigation of alternative real business cycle specifications, like Hartley et al.’s investigation of alternative data-generating processes, reminds us that no test of the real business cycle is likely on its own to provide a decisive Popperian refutation. The very fact that the models are idealized implies that the actual data alone provide at best a weak standard. More important than simply fitting the data is the relative performance of alternative models. Canova et al. and Hartley et al. push in the right direction, though not terribly far. Of course, the advocates of real business cycle models have always judged them relatively against other models in their class. Hansen’s (1985 [8]) model with indivisible labor was judged superior to his model with divisible labor. Cooley and Hansen (1989 [11]) present a real business cycle model with a monetary sector and additional monetary shocks; Christiano and Eichenbaum (1992 [10]) present one with a government sector and fiscal policy shocks. Other models have included household production (Benhabib, Rogerson, and Wright, 1991) or variable capacity utilization (Greenwood, Hercowitz, and Huffman, 1988). All of these models, however, retain the common core of the original Kydland and Prescott real business cycle model. The only substantial comparison between a real business cycle model and one with quite different principles of construction is found in Farmer’s (1993) model of an economy with increasing returns to scale and shocks to “animal spirits.” In Farmer’s model there are multiple equilibria. The economy ends up in one equilibrium or another depending upon the self-fulfilling expectations of consumers. Farmer argues that his model performs better than the real business cycle model using Kydland and Prescott’s standard of mimicking the relative correlations of actual data. He also claims that his model captures the dynamics of the economy more accurately. He estimates vector autoregressions for the actual economy and then uses the estimated equations to generate the path the economy would follow as the result of shocks to the various variables (i.e., impulse response functions). He then compares the impulse response functions of the real business cycle model and of his model with multiple equilibria to each other and to those of the © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 24 INTRODUCTION estimated VARs. He finds that the impulse responses of the real business cycle model are very different from his model and that his model is more like those from the VAR. Once again, the appeal is to aesthetic R2. Further work on standards of comparison is much to be desired.24 (iii) Testing the elements of the real business cycle model: the impulse mechanism Rather than assessing the performance of the real business cycle model directly against the data, we can ask how well its fundamental components succeed. As we noted earlier, one of two distinguishing features of the real business cycle model is that it locates the impulse to business cycles in technology shocks. The overarching question is: What evidence do we have that technology shocks are the principal impulse driving the business cycle? Before we can answer that question, however, another more basic one must be answered: What are technology shocks? This is a question that has plagued real business cycle research from the beginning (see, e.g., Summers, 1986 [5]). At the formal level, technology shocks are just the deviations of the parameter Z in the aggregate production function (e.g., equations (1.1) or (1.1') above) from its steady-state growth path: we represented these shocks earlier as z. By analogy to the microeconomic production function for individual products, one might naturally interpret z as a measure of physical technique or organizational ability. An aggregate measure should average out shocks to particular technology, so that z should measure shocks that have widespread effects across the economy. Such averaging should reduce the variability of the aggregate shocks relative to the underlying shocks to individual technologies. However, in order to make the real business cycle model match the variability of US output, the technology shocks must be relatively large and persistent: Kydland and Prescott (1982 [3]) model z as an autoregressive process with a half-life of about 14 quarters and a standard deviation of 2.9 percent of trend real per capita GDP. Our calculations for the period 1960:1 to 1993:1 are similar, yielding a standard deviation for z of 2.8 percent and for GDP per capita of 4.4 percent about trend. Although technology is improving over time, Kydland and Prescott’s assumptions about the variability of z imply that technology must sometimes regress. But as Calomiris and Hanes (1995:369–70) write: “Technological regress does not appear to correspond to any event in Western economic history since the fall of the Roman Empire.” Elsewhere they point to the large literature on the introduction and diffusion of particularly important technologies through history: even for such crucial technological developments as the steam engine, the electric motor, and the railroad, the speed of diffusion is relatively slow, so that new © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE LIMITS OF BUSINESS CYCLE RESEARCH 25 technologies take decades rather than quarters to spread through the economy. Calomiris and Hanes (1995:369): conclude that the diffusion of any one technological innovation could not increase aggregate productivity by more than a trivial amount from one year to the next. If no single innovation can make much of a difference, it seems extremely unlikely that the aggregate rate of improvement could vary exogenously over cyclical frequencies to an important degree. In the face of such objections, proponents of real business cycle models have broadened the scope of technology to include “changes in the legal and regulatory system within a country” (Hansen and Prescott, 1993 [26]: 281). Fair enough; such changes may be important to the economy and may plausibly be negative; but are they likely to justify quarter-to-quarter variation in z of the required amount? Furthermore, as Calomiris and Hanes (1995:370) point out, regulatory and legal intervention in the US economy was substantially smaller before World War I when business cycles themselves were more variable.25 Debates over the size and frequency of technology shocks are difficult to resolve because the shocks are not directly observable. Real business cycle models have largely adopted the biblical criterion “by their fruits ye shall know them” and used the Solow residual (equation (1.7) above) as a proxy for technology shocks. The Solow residual attributes to technology any change in output that cannot be explained by changes in factor inputs. Jorgenson and Griliches (1967) and Griliches (1996 [29]) point out that the Solow residual measures more than underlying technological change (a fact recognized by Solow, 1957 [27]: 312, himself), picking up among other things variability in capital utilization and labor hoarding.26 Summers (1986 [5]) and Mankiw (1989 [28]) reiterate these points in the context of real business cycle models. Hall (1986, 1990) notes that calibrating the parameters of the Cobb-Douglas production function (equation (1.1')), and (1-), as the shares of labor and capital in output in the calculation of the Solow residual (as in equation (1.7)) requires the assumption of perfect competition so that firms and workers are paid their marginal products and factor shares exactly exhaust output. But if firms have market power so that price exceeds marginal cost, factor shares will no longer coincide with the technological parameters and (1-), and z will reflect variations in markups across the business cycle as well as true technology shocks. Hall (1990) also demonstrates that if there are increasing returns to scale, the Solow residual will move with things other than pure technology shocks. Jorgenson, Griliches and Hall conclude that the Solow residual captures a great deal besides technology. Hartley (1994) provides evidence that the © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 26 INTRODUCTION Solow residual may not reliably capture even genuine technology shocks. The evidence is found in simulated economies constructed using Hansen and Sargent’s (1990) flexible, dynamic linear-quadratic equilibrium macromodel. This model permits a richer specification of the underlying production technology than typical of real business cycle models: there are multiple sectors, including intermediate and final goods, and parameters representing multiple aspects of the production process. Hartley was able to generate series for output, capital, and labor based on shocks to specific parts of the production process. Because these were simulations, he could be assured that the variability in these series reflected only technology shocks and not market power, labor hoarding, and the like. He then calculated Solow residuals from the simulated series using equation (1.7) and asked whether these accurately reflected the size and direction of the underlying true technology shocks. For a wide range of plausible parameters, Hartley found an extremely low correlation between his controlled technology shocks and the calculated Solow residuals. Often the correlation was not even positive. The failure of the Solow residual to capture the underlying process accurately appears to reflect the fact that the Cobb-Douglas production function, implicit in the calculation of Solow residuals, is a poor approximation to the rich production details of Hansen and Sargent’s model: the Solow residuals largely reflect specification error, rather than technological change on a quarter-by-quarter horizon. Hansen and Sargent’s model is rich relative to the typical idealized real business cycle model, but is itself an extreme idealization of the real production process. Hartley’s simulation results, a fortiori, call the Solow residual into question as a measure of actual technology shocks. (iv) Testing the elements of the real business cycle model: the propagation mechanism The propagation mechanism of a business cycle model should transmit and amplify the impulses to the various cyclical aggregates. Together with the shocks themselves it should account for the pattern of fluctuations in each series and for their comovements. Real output is generally taken as the marker series for the business cycle. The balance of evidence is that real business cycle models add relatively little to the pattern of fluctuations in real output beyond what is implicit in the technology shocks themselves. Watson (1993 [17]) uses spectral analysis to decompose the power of the real business cycle model to match movements in output at different frequencies or (equivalently) time horizons. He finds that the spectral power of the real business cycle model is high at low frequencies (corresponding to trend or long-term growth behavior), but low at business cycle frequencies (approximately two to eight years). Cogley and Nason (1995b [30]) compare the dynamic pattern of the technology shocks fed © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE LIMITS OF BUSINESS CYCLE RESEARCH 27 into the real business cycle model with the predicted time series for output generated by the model. Again, they find that it is the dynamic properties of the exogenous inputs that determine the properties of the output, with the model itself contributing almost nothing. In one sense, these results should not be surprising: the Solow growth model, the foundational model of the real business cycle model, was originally meant to capture secular change. It is bold to conjecture that, unaltered, it would also model the business cycle. What is more surprising is that it took relatively long to document its low value-added with respect to business cycles. Part of the reason that the real business cycle model has appeared to do well is that its proponents—sometimes for good methodological reasons— have relied on standards of assessment that are not particularly discriminating and have failed to develop more discriminating ones (see section II (ii) above). Part of the reason has to do with the standard practices of real business cycle models with respect to handling data. The real business cycle model predicts values for output, consumption, investment, and other time series expressed as deviations from the steady state. In order to compare these with actual data, an estimate of the steady state must be removed from these variables, which typically are trending. The Solow growth model suggests that all these variables should grow at rates related to the steady-state growth rate. Unfortunately, that is not observable. In practice, real business cycle models follow one of two strategies to generate detrended data. They sometimes remove a constant exponential trend, which is linear in the logarithm of the series, and so is known as linear detrending (e.g., King, Plosser and Rebelo, 1988 [7]). This would be precisely correct if the growth model were in fact true and the rate of growth of the labor force (n) and of technology () were constant over time, so that the steady-state growth rate (n+) were also constant over time. But there is no reason to think that this is so. An alternative strategy is to use a slowly varying trend that effectively allows the steady-state growth rate to be variable. This is the most popular option and it is typically implemented using the Hodrick-Prescott (HP) filter (Hodrick and Prescott, 1997 [31]).27 The HP filter is a nonlinear regression technique that acts like a two-side moving average. As we noted, and as Prescott (1986a [4]) asserts, one should in principle model growth and cycles jointly (see also Kydland and Prescott, 1996 [13]). In practice, however, real business cycle models express the interrelationships of data as deviations from the steady state. So, in effect, the HP filter provides an atheoretical estimate of the steady-state growth path. Harvey and Jaeger (1993 [32]) analyze the usefulness of the HP filter in accomplishing this task. They compare the cyclical component for output generated from an HP filter to that from a structural time-series model in which the trend and the cyclical component are estimated jointly. (This is closer to what Kydland and Prescott advocate in principle than to what © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 28 INTRODUCTION they actually practice.) For US GDP, both detrending methods produce similar cyclical components. Harvey and Jaeger, however, demonstrate that the H P filter is wildly different from the structural time-series model for several other countries. This underscores the previously cited finding of Hartley et al. (1997 [24]) that the real business cycle model matches US data—at least on the standard of aesthetic R2 typically employed by real business cycle modelers—but not artificial data of a more “European” character. Harvey and Jaeger also show that the HP filter and the structural timeseries model differ substantially when applied to other US time series— particularly in the case of U.S. prices and the monetary base. Given Kydland and Prescott’s impassioned attack on the “monetary myths” of the business cycle, it is obviously critical to know whether the facts about money and prices are independent of the filtering process. Furthermore, Harvey and Jaeger demonstrate that in small samples the HP filter can induce apparent cyclical fluctuations and apparent correlations between series even when the prefiltered series are independent and serially uncorrelated. As they point out, these results are in the same spirit as Slutsky’s and Yule’s analyses of spurious cyclical behavior (Yule (1926), Slutsky ([1927] 1937); more recently, see Nelson and Kang, 1981). This phenomenon has been long known if not fully appreciated. Simon Kuznets, for example, “discovered” long cycles in US data that had first been transformed through two separate moving averages and first differencing. It can be shown that purely random data subjected to such transformations present precisely the same twenty-year cycles that Kuznets reported: they are nothing but an artifact of the filtering (see Sargent, 1979:249–51). The analogy between the HP filter and Kuznets’s transformations is close because the HP filter acts as a type of two-sided moving average. Cogley and Nason (1995a [33]) reinforce Harvey and Jaeger’s analysis; they demonstrate that prefiltered data do not generate cycles in a real business cycle model, while HP filtered data do. Furthermore, when the input data are highly serially correlated (a correlation coefficient of unity, or nearly so, between current and lagged values of the variable: i.e., a unit root or near unit root), the HP filter not only generates spurious cycles but also strongly increases the correlation among the predicted values for output, consumption, investment, hours of work, and other predicted values from the real business cycle model. The model itself— that is, the propagation mechanism—does little of the work in generating the cyclical behavior; the HP filter does the lion’s share. On the one hand, the use of the HP filter calls into question the very facts of the business cycle. Kydland and Prescott (1990 [20]) document the intercorrelations among HP filtered time series. These correlations are held by real business cycle modelers to provide strong prima facie support for the real business cycle model (Kydland and Prescott’s (1990 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE LIMITS OF BUSINESS CYCLE RESEARCH 29 [20]) subtitle to their paper is “Real Facts and a Monetary Myth”). For example, they show that the correlation between HP-filtered real GDP and HP-filtered prices is -0.50, and claim that this contradicts the prediction of Keynesian models that prices are procyclical. Harvey and Jaeger (1993 [32]) not only show that the HP filter can induce such correlations, but they also show that it adds statistical noise, so that a genuine correlation would, in a sample size of one hundred, have to exceed 0.40 before we could be sure that it was statistically significant at the standard 5 percent critical value. If such correlations are really artifacts of a filtering procedure, with no particular grounding in the economics of the business cycle, then the support of the “real facts” for the real business cycle model is substantially weakened. Prescott (1986a [4]: 10) wrote: “If the business cycle facts were sensitive to the detrending procedure used, there would be a problem. But the key facts are not sensitive to the procedure if the trend curve is smooth.” The weight of evidence since Prescott wrote this suggests that he is incorrect: the facts are sensitive to the type of filtering that defines the trend. On the other hand, while there is good reason to find some way to detrend the technology-shock series used as an input into the real business cycle model, it is also standard practice to HP filter the predicted series generated by the real business cycle model before checking their intercorrelations and comparing them to the HP filtered actual data. Harvey and Jaeger’s and Cogley and Nason’s analyses suggest that this practice raises the correlations among these series artificially. Kydland and Prescott (1996 [13]: 76–77 n) defend the use of the HP filter against critics who have argued that it induces spurious cycles by stating that deviations from trends defined by the HP filter “measure nothing,” but instead are “nothing more than well-defined statistics”; and, since “business cycle theory treats growth and cycles as being integrated, not as a sum of two components driven by different factors… talking about the resulting statistics as imposing spurious cycles makes no sense.” The logic of Kydland and Prescott’s position escapes us. It is true that real business cycle theory treats the business cycle as the equilibrium adjustments of a neoclassical growth model subject to technology shocks. But, as we have previously noted, the real business cycle model does not, in practice, model the steady state. The HP filter is an atheoretical method of extracting it prior to the economic modeling of the deviations from the steady state. The implicit assumption is that the extracted trend is a good approximation of the steady state, for which no evidence is offered. This does not say that the steady state could not be modeled jointly with the deviations in principle. That it is not actually modeled jointly in practice, however, means that the objection to the HP filter raised by many critics remains cogent. The work of Harvey and Jaeger and Cogley and Nason (see also Canova, 1991), which Kydland and Prescott wish to dismiss, © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 30 INTRODUCTION demonstrates that the choice of which ad hoc method is used to extract the balanced-growth path greatly affects the stochastic properties of the modeled variables and their relationships with the actual data. One way of reading Watson (1993 [17]) and Cogley and Nason (1995a [30]) is that, while a model driven by a technology shocks fits output well, it is the technology shocks not the model which are responsible for that fact. The picture painted is one of the real business cycle model as a slightly wobbly transmission belt converting the time-series characteristics of the technology shocks into the model’s predictions for real output. But in the end there is a good fit between the model and real output. King (1995) and Hoover (1997) suggest that if the Solow residual is taken as the proxy for technology shocks, then this success is an illusion. Despite having rejected in earlier work the relevance of direct comparisons to historical data, real business cycle models have recently made precisely such comparisons.28 Hansen and Prescott (1993 [26]) ask whether technology shocks can explain the 1990–91 recession in the United States, while Cooley (1995a) asks whether they can explain the “Volcker recessions” of 1980–82. In each case, the predictions of a real business cycle model are compared directly to the historical path of real output.29 Again the standard is one of aesthetic R2, and the pitfalls of this standard are easily seen Hansen and Prescott’s (1993 [26]: 285) figure 4 (see p. 538 below). Hansen and Prescott cite the fact that the output predicted from their real business cycle model tracks actual output as favorable evidence for its explanatory power. In particular, they note that the model catches the fall in output in 1990–91. But look more closely. Actual GNP peaks in the first quarter of 1990, while model GNP peaks in the fourth quarter; actual GNP bottoms out in the first quarter of 1991, while model GNP bottoms out in the second quarter. Furthermore, the model predicts two earlier absolute falls in GNP, while, in fact, there are no other recessions in the data. One of these predicted falls is actually on a larger scale than the genuine recession of 1990–91: the model shows that GNP peaks in the first quarter of 1986 and falls 2.3 percent to a trough in the fourth quarter of 1986, where in reality GNP rose the entire time. The actual fall in GNP in the 1990–91 recession is only 1.6 percent. The difficulties of using aesthetic R2 to one side, these graphical measures or their more statistical counterparts (e.g., see Smith and Zin, 1997) offer no support for the real business cycle model. To see the difficulty, consider a simple version of a real business cycle model in which we abstract from time trends. Initially, let labor be supplied inelastically. Capital is inherited from the previous period. The Solow residual (zt) can be calculated in loglinear form: (1.8) The log-linear version of the production function is given by © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE LIMITS OF BUSINESS CYCLE RESEARCH 31 (1.9) where the s subscripts refer to variables determined in the model. From the inelastic supply of labor we know that Ls=Lt. Substituting this fact and the Solow residual into equation (1.9), (1.10) or (1.10´) Our model fits perfectly; the correlation between predicted and actual output is unity. Does anyone believe that this is a demonstration of its goodness? Real business cycle models do not fit perfectly, as this little exercise suggests. The reason is that the inputs to their production function do not recapitulate the capital and labor measures used to calculate the Solow residual. In particular, the labor input (Ls) is determined by other features of the model—in fact, by features that are considered the most characteristic of real business cycle models, such as intertemporal substitutability of labor and optimal investment and consumption planning.30 So, it is natural to relax our assumption of an inelastic labor supply. Then equation (1.10) becomes (1.11) How well predicted output fits actual output is seen to depend on how well predicted labor fits actual labor. Still, there is an artifactual element to the correlation between predicted and actual output. Notice that the share parameter is not given in nature, but is a modeling choice. If is calibrated to be close to zero, then the predicted and actual output are again nearly perfectly correlated. Now, it might be objected that we know is not close to zero but close to 0.69 (Hansen and Prescott’s (1993 [26]) assumption). We agree. But information about the true size of comes from the calibrator’s supply of exogenous information and has nothing to do with the fit of the model to historical data or with traditional econometrics. It underscores Kydland and Prescott’s advocacy of external sources of information to pin down free parameters. We must not forget that whether is zero, 0.69, or one, actual output shows up on the righthand side of equation (1.11) only because we put it there in the construction of the Solow residual, not because the model generated it by closely matching the structure of the economy.31 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 32 INTRODUCTION Of course, it would be a marvelous testament to the success of the model if the right-hand side of equation (1.11) turned out to be very nearly log(Yt). That would occur because the model’s predicted labor was very nearly the actual labor. The true measure of such success is not found indirectly in the comparison of Ys to Yt, but in the direct comparison of Ls to Lt. Even a test based on modeling labor conditional on the Solow residual is likely to suffer from spurious success at fitting historical labor, since the Solow residual also contains current labor information by construction. Truly revealing tests of the success of the real business cycle model at capturing the true propagation mechanism based on comparisons of the predictions of the model against historical time series should then concentrate on those series (e.g., consumption) the current values of which play no part in the construction of measures of the technology shocks. To give up the comparison of historical and predicted output (or labor) is not to give up the comparison of historical and predicted data altogether. One might ask different questions of the model: for example, if one knew actual output and capital, what would the model imply that consumption and labor would have to be? These conditional predictions are measures of consumption and labor that are uncorrupted by actual labor in their construction. Historical comparisons on these dimensions would be useful tests of the model: a close fit would then be a genuine accomplishment of the real business cycle model, and not an artifact of the construction of the Solow residual.32 We know of no work to date that has systematically investigated the propagation mechanism of the real business cycle model in this manner independently of the Solow residual. III REFUTATION? The history of real business cycle models illustrates a fact well known to philosophers and historians of science: It is rare for a conjecture—however bold—to be refuted simpliciter on the basis of a single experiment or a single observation, as in Popper’s ideal case. Accumulated evidence may, nonetheless, render the intellectual cost of persisting in a particular conjecture (model or theory) higher than the cost of abandoning or modifying it. To some extent, it does not appear to be controversial that the evidence weighs against the real business cycle program narrowly construed. Even the best-known advocates of real business cycle models have tended to move away from models of perfectly competitive representative agents driven by technology shocks only (see n. 1). While these models are direct descendants of the real business cycle model and remain in the broader class of equilibrium business cycle models, they represent an abandonment of the strongest form of the real business cycle conjecture. The balance of the evidence presented here is that they are © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE LIMITS OF BUSINESS CYCLE RESEARCH 33 right to do so. Although there can be no objection to investigating just how far these new models can be pushed, there is little in the evidence with respect to the narrower real business cycle conjecture that would warrant much optimism about their success. The case against the real business cycle conjecture has several parts. First, the propagation mechanism (i.e., the Solow growth model), while it provides, to a first approximation, a reasonable account of long-term growth, has virtually no value added with respect to business cycles. The growth model will transmit fluctuations at business cycle frequencies from impulses that are already cyclical, but it will not generate them from noncyclical impulses. The putative impulse mechanism is the fluctuation of technology. In the model itself this amounts to shifts in a disembodied parameter (Z). The proponents of real business cycle models have given very little account of what features of the world might correspond to Z and fluctuate in the way needed to produce business cycles. Z is an unexplained residual in every sense of the word: it is whatever it has to be to make the real business cycle model behave in an appropriate manner, and it cannot be independently observed. If measured as the Solow residual, “technology” means whatever bit of output that cannot be accounted for by capital and labor inputs. Using this residual output as an impulse cannot yield predicted values for output that provide a logically sound independent comparison between the model and the actual data on the dimension of output. While valid comparisons might be made on other dimensions, the actual evidence in favor of real business cycles is weak in the sense that it does not provide discriminating tests: alternative models do as good a job in mimicking the data on the usual aesthetic standards as does the real business cycle model. Both the facts to be explained and the ability of the models to match those facts are themselves frequently distorted by the common data-handling techniques (particularly the HP filter). These data problems, combined with the fact that the highly idealized nature of the real business cycle models limits the ambitions that their advocates have for matching the actual data, insulate the model from decisive refutation, but equally well undercut the role of empirical evidence in lending positive support to them. The real business cycle model has for fifteen years dominated the agenda of business cycle research. On balance, however, there is little convincing empirical evidence that favors it over alternative models. To its advocates, the paucity of evidence may not be of too much concern, for Kydland and Prescott (1991 [12]: 171) argue that the confidence one places in a model to answer economic question cannot “be resolved by computing some measure of how well the model economy mimics historical data…. The degree of confidence in the answer depends on the confidence that is placed in the economic theory being used.” But anyone who believes that © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 34 INTRODUCTION theories must be warranted by evidence has little reason to date to place much confidence in real business cycle models. NOTES 1 As will become clear below, our focus is on real business cycles narrowly construed as perfectly competitive representative agent models driven by real shocks. A number of recent developments have extended models with roots in Kydland and Prescott (1982 [3]) to include monetary factors, limited heterogeneity among agents, non-Walrasian features, and imperfect competition. These models are ably surveyed in chapters 7–9 of Cooley (1995b). One way to view this literature is as a constructive response to some of the difficulties with the narrow real business cycle model that we evaluate. 2 Lucas (1978:242). 3 Friedman’s monetarist model is distinguished from Lucas’s new classical monetary model in that Friedman imagines that people can be systematically mistaken about the true state of real wages for relatively long periods, while Lucas argues that people have rational expectations (i.e., they make only unsystematic mistakes) and, therefore, correct their judgments about the real wage quickly. 4 Lucas (in Snowden, Vane and Wynarczyk, 1994:222) accepts that his previous characterization of the Austrians as precursors to new classical business cycle theory was incorrect. 5 Equation (1.1) is a snapshot of the economy at a particular time. In fact, variables in the model are growing. We could indicate this with subscripts indexing the relevant time, but this would simply clutter the notation unnecessarily. 6 There is a large literature on endogenous growth models (see, e.g., the symposium in the Journal of Economic Perspectives, 1994). 7 Cooley and Prescott (1995, sec. 4) discusses the issues related to establishing an appropriate correspondence between the real business cycle model and the national accounts to permit the calibration of the model. 8 It is actually a debated question whether microeconomic studies do in fact provide the necessary parameters. Prescott (1986a [4]: 14) cites Lucas’s (1980:712) argument that we have “a wealth of inexpensively available data” of this sort. However, Hansen and Heckman (1996 [14]: 93–94) argue that in this regard Prescott is wrong. As evidence they point to Shoven and Whalley (1992:105), who rather candidly admit that “it is surprising how sparse (and sometimes contradictory) the literature is on some key elasticity values. And although this procedure might sound straightforward, it is often exceedingly difficult because each study is different from every other.” (Cf. the debate between Summers (1986 [5]) and Prescott (1986b [6]) about whether the parameters used in Prescott (1986a [4]) are the appropriate ones.) 9 Details on how to solve these sorts of models are provided in Chapter 2. 10 Despite this argument, Lucas’s view of real business cycle models is rather favorable. See, e.g., the discussion in Manuelli and Sargent (1988). 11 Although we refer to z as “the technology shock,” this terminology is not universal. Generally, z will be a persistent process; for example, zt=zt+εt, with >0 and εt an independent, identically distributed random variable. Some economists identify εt as “the technology shock.” Similarly, some economists identify zt rather than Zt as the “Solow residual.” © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE LIMITS OF BUSINESS CYCLE RESEARCH 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 35 These reasons are elaborated in Hartley (1997). The classic reference is Gorman (1953); the literature is summarized in Stoker (1993). Coleman (1990) calls this the micro-to-macro transition and provides an extensive and illuminating discussion about what is involved in making this transition properly. This seems to be Friedman’s (1997:210) point when he criticizes Kydland and Prescott’s (1996 [13]) standards of empirical evaluation for calibrated models, saying, “There is a world of difference between mimicking and explaining, between ‘can or may’ and ‘does’.” The official definition of a business cycle used by the National Bureau of Economic Research in the United States emphasizes that recessions are unfavorable movements across a breadth of economic aggregates; see Geoffrey Moore quoted in Hall (1981). Prescott (1986a [4]: 10) argues that the noun “business cycle” should be avoided as it encourages people to believe incorrectly that there is an entity to be explained, independently of economic growth, which is characterized by a deterministic cycle. Instead, “business cycle” should be used as an adjective, as in “business cycle phenomena,” that points to the volatility and comovements of various economic series. Lucas (1987, sec. V) acknowledges the difference between defining the business cycle, as common usage does, as recurrent fluctuations in unemployment and, as he and real business cycle models typically do, as equilibrium comovements. He recognizes that, to the extent that one is interested in questions of unemployment, models that aim to explain the comovements alone are silent on an important question—although he argues that this is a limitation, not a fault. Harvey and Jaeger (1993) present evidence that the HP filter which is used to detrend the series analyzed by Kydland and Prescott (1990 [20]) distorts the correlations among them, suggesting that the “facts” might be artifacts of the statistical processing after all (see section II (v) below). Farmer (1993) is an exception, see section II (ii) below. Both the term “stylized facts” and the facts themselves are due to Kaldor (1961). Also see Kaldor (1957) in which the facts themselves are discussed with the name “stylized.” Manuelli and Sargent (1988) criticize the real business cycle literature for backing away from following procedures along these lines. Kydland and Prescott (1990 [20]) and Burns and Mitchell (1946). See Fair (1990) for a description of the model. A strategy for the assessment of idealized models is discussed in Hoover (1994). This claim is controversial. Romer (1986a, b; 1989) argues that postwar business cycles are not substantially less variable than those of the nineteenth century. Weir (1986) and Balke and Gordon (1989) challenge Romer’s revisionism. The debate is updated and assessed in Siegler (1997), which, on the basis of better estimates of nineteenth-century GNP, supports the traditional view that modern business cycles are in fact smoother than those of the nineteenth century. Solow (1956 [27]: 314, 320) explicitly observes that idle capacity biases the measure and that the measure hinges on the assumption of factors being paid their marginal products, but that a similar measure could be created for monopolistic factor markets. Looking back over thirty years later, Solow (1990:225) argues that he never intended the Solow residual as a suitable measure of anything but © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 36 INTRODUCTION the trend in technology: “the year-to-year behavior of the residual could be governed by all sorts of ‘technologically’ irrelevant short-run forces. I still think that….” 27 The HP filter is defined as follows: Let xt=x-t + x^t where x-t denotes the trend component and x^t denotes the deviation from trend. The HP filter chooses this decomposition to solve the following problem: 28 29 30 31 32 T is the number of observations and λ is a parameter that controls the amount of smoothness in the series: if λ=0, then the smooth series is identical to the original series; if λ=∞, the smoothed series is just a linear trend. Hodrick and Prescott use a value of λ=1600 for quarterly data on the ground that this replicates the curve a business cycle analyst might fit freehand to the data. With no better justification than this, λ=1600 has become the standard choice for the smoothing parameter in the real business cycle literature. For example, Prescott (1986a [4]: 16) argues against comparing the model output to the path of actual US output. Christiano (1988) seems to have been the first real business cycle modeler to employ such a test. Additionally, independently detrending the Solow residual and the other inputs to the production function may introduce discrepancies between actual and model-generated data. Hoover and Salyer (1996) provide simulation evidence that the Solow residual does not convey useful information about technology shocks, and that the apparent success of real business cycle models in matching historical data for output is wholly an artifact of the use of current output in the construction of the Solow residual. If, for instance, we were to condition on actual output, inherited capital and the government expenditure shock, then we could back out another measure of zS. But, given that we have nothing independent to compare it with, the more interesting point is that we can back out some other series, say, labor conditional on actual output and capital, which can then be compared to its actual counterpart. REFERENCES Adelman, Irma and Frank L.Adelman (1959) “The Dynamic Properties of the Klein-Goldberger Model,” Econometrica, 27(4): 596–625. Altu, Sumru (1989) “Time-to-Build and Aggregate Fluctuations: Some New Evidence,” International Economic Review, 30(4), November: 889–920, reprinted here in Chapter 21. Backus, David K. and Patrick J.Kehoe (1992) “International Evidence on the Historical Properties of Business Cycles,” American Economic Review, 82(4): 864–88. Balke, Nathan S. and Robert J.Gordon (1989) “The Estimation of Prewar Gross National Product: Methodology and New Evidence,” Journal of Political Economy, 97(1): 38–92. Benhabib, Jess, Richard Rogerson, and Randall Wright (1991) “Homework in Macroeconomics: Household Production and Aggregate Fluctuations,” Journal of Political Economy, 99(6), December: 1166–87. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE LIMITS OF BUSINESS CYCLE RESEARCH 37 Bryant, John (1983) “A Simple Rational Expectations Keynes-type Model,” Quarterly Journal of Economics, 98(3), August: 525–28. Burns, Arthur F. and Wesley C.Mitchell (1946) Measuring Business Cycles, New York: National Bureau of Economic Research. Calomiris, Charles W. and Christopher Hanes (1995) “Historical Macroeconomics and Macroeconomic History,” in Kevin D.Hoover (ed.) Macroeconometrics: Developments, Tensions, and Prospects, Dordrecht: Kluwer: 351–416. Canova, Fabio (1991) “Detrending and Business Cycle Facts,” unpublished typescript, Department of Economics, European University Institute, Florence, Italy, July 20. Canova, Fabio, M.Finn, and A.R.Pagan (1994) “Evaluating a Real Business Cycle Model,” in C.Hargreaves (ed.) Nonstationary Time Series Analysis and Cointegration, Oxford: Oxford University Press: 225–55, reprinted here in Chapter 22. Christiano, Lawrence J. (1988) “Why Does Inventory Investment Fluctuate So Much?” Journal of Monetary Economics, 21(2): 247–80. Christiano, Lawrence J. and Martin Eichenbaum (1992) “Current Real Business Cycle Theories and Aggregate Labor Market Fluctuations,” American Economic Review, 82(3), June: 430–50, reprinted here in Chapter 10. Cogley, Timothy and James M.Nason (1995a) “Effects of the Hodrick-Prescott Filter on Trend and Difference Stationary Time Series: Implications for Business Cycle Research,” Journal of Economic Dynamics and Control, 19(1– 2), January-February: 253–78, reprinted here in Chapter 33. Cogley, Timothy and James M.Nason (1995b) “Output Dynamics in Real Business Cycle Models,” American Economic Review, 85(3), June: 492–511, reprinted here in Chapter 30. Coleman, James S. (1990) Foundations of Social Theory, Cambridge, Mass.: The Belknap Press. Cooley, Thomas F. (1995a) “Contribution to a Conference Panel Discussion: What Do We Know about How Monetary Policy Affects the Economy?” paper given at the 19th Annual Economic Policy Conference, Federal Reserve Bank of St Louis. Federal Reserve Bank of St Louis Review, 77(3): 131–37. Cooley, Thomas F. (ed.) (1995b) Frontiers of Business Cycle Research, Princeton: Princeton University Press. Cooley, Thomas F. and Gary D.Hansen (1989) “The Inflation Tax in a Real Business Cycle Model,” American Economic Review, 79(4), September: 733– 48, reprinted here in Chapter 11. Cooley, Thomas F. and Edward C.Prescott (1995) “Economic Growth and Business Cycles,” in Cooley (1995b): 1–38. Cooper, Russell and Andrew John (1988) “Coordinating Coordination Failures,” Quarterly Journal of Economics, 103(3), August: 441–64. Debreu, Gerard (1974) “Excess Demand Functions,” Journal of Mathematical Economics, 1(1), March: 15–23. Eichenbaum, Martin (1991) “Real Business Cycle Theory: Wisdom or Whimsy?” Journal of Economic Dynamics and Control, 15(4), October: 607– 26, reprinted here in Chapter 25. Fair, Ray C. (1990) Fairmodel: User’s Guide and Intermediate Workbook, Southborough, Mass.: Macro Incorporated. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 38 INTRODUCTION Farmer, Roger E.A. (1993) The Macroeconomics of Self-fulfilling Prophecies, Cambridge, Mass.: MIT Press. Friedman, Milton (1957) A Theory of the Consumption Function, Princeton: Princeton University Press. Friedman, Milton (1968) “The Role of Monetary Policy,” American Economic Review, 58(1): 1–17. Friedman, Milton (1997) “Computational Experiments,” Journal of Economic Perspectives, 11(1), Winter: 209–10. Frisch, Ragnar (1933) “Propagation Problems and Impulse Response Problems in Dynamic Economics,” in Economic Essays in Honour of Gustav Cassel: October 20, 1933, London: George Allen & Unwin. Gorman, W.M. (1953) “Community Preference Fields,” Econometrica, 21(1): 63–80. Greenwood, Jeremy, Zvi Hercowitz, and Gregory W.Huffman (1988) “Investment, Capacity Utilization and the Real Business Cycle,” American Economic Review, 78(3), June: 402–17. Gregory, Allan W. and Gregor W.Smith (1990) “Calibration as Estimation,” Econometric Reviews, 9(1): 57–89. Gregory, Allan W. and Gregor W.Smith (1991) “Calibration as Testing: Inference in Simulated Macroeconomic Models,” Journal of Business and Economic Statistics, 9(3), July: 297–303, reprinted here in Chapter 16. Griliches, Zvi (1996) “The Discovery of the Residual: A Historical Note,” Journal of Economic Literature, 34(3): 1324–30, reprinted here in Chapter 29. Hall, Robert E. (1981) “Just a Peak and Trough,” National Bureau of Economic Research (downloaded from World Wide Web site: nber.harvard.edu). Hall, Robert E. (1986) “Market Structure and Macroeconomics Fluctuations,” Brookings Papers on Economic Activity, no. 2:265–338. Hall, Robert E. (1990) “Invariance Properties of Solow’s Productivity Residual,” in Peter Diamond (ed.) Growth/Productivity/Unemployment: Essays to Celebrate Bob Solow’s Birthday, Cambridge, Mass.: MIT Press: 71–112. Hansen, Gary D. (1985) “Indivisible Labor and the Business Cycle,” Journal of Monetary Economics, 16(3), November: 309–28, reprinted here in Chapter 8. Hansen, Gary D. and Edward C.Prescott (1993) “Did Technology Shocks Cause the 1990–1991 Recession?” American Economic Review, 83(2), May: 280–86, reprinted here in Chapter 26. Hansen, Gary D. and Thomas J.Sargent (1988) “Straight Time and Overtime in Equilibrium,” Journal of Monetary Economics, 21(2), March: 281–308. Hansen, Gary D. and Randall Wright (1992) “The Labor Market in Real Business Cycle Theory,” Federal Reserve Bank of Minneapolis Quarterly Review, 16(2), Spring: 2–12, reprinted here in Chapter 9. Hansen, Lars Peter and James J.Heckman (1996) “The Empirical Foundations of Calibration,” Journal of Economic Perspectives, 10(1), Winter: 87–104, reprinted here in Chapter 14. Hansen, Lars Peter and Thomas J.Sargent (1980) “Formulating and Estimating Dynamic Linear Rational Expectations Models,” Journal of Economic Dynamics and Control, 2(1), February, reprinted in Robert E.Lucas, Jr. and Thomas J. Sargent (eds) (1981) Rational Expectations and Econometric Practice, London: George Allen & Unwin: 91–125. Hansen, Lars Peter and Thomas J.Sargent (1990) “Recursive Linear Models of Dynamic Economies,” National Bureau of Economic Research, Working © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE LIMITS OF BUSINESS CYCLE RESEARCH 39 Paper Series, no. 3479. Hartley, James E. (1994) “Technology in Macroeconomic Models,” doctoral dissertation, University of California, Davis. Hartley, James E. (1997) The Representative Agent in Macroeconomics, London: Routledge. Hartley, James E., Kevin D.Salyer, and Steven M.Sheffrin (1997) “Calibration and Real Business Cycle Models: An Unorthodox Experiment,” Journal of Macroeconomics, 19(1): 1–17, reprinted here in Chapter 24. Harvey, A.C. and A.Jaeger (1993) “Detrending, Stylized Facts and the Business Cycle,” Journal of Applied Econometrics, 8(3): 231–47, reprinted here in Chapter 32. Hayek, Friedrich A. von (1933) Monetary Theory and the Trade Cycle, trans. Nicholas Kaldor and H.M.Croome, London: Jonathan Cape. Hayek, Friedrich A. von (1935) Prices and Production, 2nd ed., London: Routledge & Kegan Paul. Hodrick, Robert J. and Edward C.Prescott (1997) “Postwar US Business Cycles: An Empirical Investigation,” Journal of Money, Credit, and Banking, 29(1), February: 1–16, reprinted here in Chapter 31. Hoover, Kevin D. (1988) The New Classical Economics: A Sceptical Inquiry, Oxford: Blackwell. Hoover, Kevin D. (1994) “Six Queries About Idealization in an Empirical Context,” Poznan Studies in the Philosophy of Science and the Humanities, 38:43– 53. Hoover, Kevin D. (1995) “Facts and Artifacts: Calibration and the Empirical Assessment of Real-Business-Cycle Models,” Oxford Economic Papers, 47(1), March: 24–44, reprinted here in Chapter 15. Hoover, Kevin D. (1997) “Comments on Smith and Zin’s ‘Real Business Cycle Realizations, 1925–1995’,” Carnegie-Rochester Series in Public Policy, forthcoming. Hoover, Kevin D. and Kevin D.Salyer (1996) “Technology Shocks or Colored Noise? Why Real Business Cycle Models Cannot Explain Actual Business Cycles,” unpublished manuscript. Hume, David ([1739] 1888) A Treatise of Human Nature, Oxford: Clarendon Press. Ingrao, Bruna and Giorgio Israel (1990) The Invisible Hand: Economic Equilibrium in the History of Science, trans. Ian McGilvray, Cambridge, Mass.: MIT Press. Jorgenson, Dale W. and Griliches, Zvi (1967) “The Explanation of Productivity Change,” Review of Economic Studies, 34(3), July: 249–83. Journal of Economic Perspectives (1994) Symposium on “New Growth Theory,” 8(1): 3–72. Kaldor, Nicholas (1957) “A Model of Economic Growth,” Economic Journal, 67(4), December: 591–624. Kaldor, Nicholas (1961) “Capital Accumulation and Economic Growth,” in Friedrich A.Lutz and Douglas C.Hague (eds) The Theory of Capital: Proceedings of a Conference Held by the International Economics Association, London: Macmillan. King, Robert G. (1995) “Quantitative Theory and Econometrics,” Federal Reserve Bank of Richmond Economic Quarterly, 81(3), Summer: 53–105. King, Robert G. and Charles I.Plosser (1989) “Real Business Cycles and the Test of the Adelmans,” Journal of Monetary Economics, 33(2), April: 405–38, reprinted here in Chapter 23. King, Robert G., Charles I.Plosser and Sergio T.Rebelo (1988) “Production, © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 40 INTRODUCTION Growth, and Business Cycles I: The Basic Neoclassical Model,” Journal of Monetary Economics, 21(2), March: 195–232, reprinted here in Chapter 7. Kirman, Alan P. (1989) “The Intrinsic Limits of Modern Economic Theory: The Emperor Has No Clothes,” The Economic Journal, 99 (Conference): 126–39. Kirman, Alan P. (1992) “Whom or What Does the Representative Individual Represent?” Journal of Economic Perspectives, 6(2), Spring: 117–36. Kydland, Finn E. and Edward C.Prescott (1982) “Time to Build and Aggregate Fluctuations,” Econometrica, 50(6), November: 1345–69, reprinted here in Chapter 3. Kydland, Finn E. and Edward C.Prescott (1988) “The Workweek of Capital and Its Cyclical Implications,” Journal of Monetary Economics, 21(2): 343– 60. Kydland, Finn E. and Edward C.Prescott (1990) “Business Cycles: Real Facts and a Monetary Myth,” Federal Reserve Bank of Minneapolis Quarterly Review, 14(2), Spring: 3–18, reprinted here in Chapter 20. Kydland, Finn E. and Edward C.Prescott (1991) “The Econometrics of the General Equilibrium Approach to Business Cycles,” Scandinavian Journal of Economics, 93(2): 161–78, reprinted here in Chapter 12. Kydland, Finn E. and Edward C.Prescott (1996) “The Computational Experiment: An Econometric Tool,” Journal of Economic Perspectives, 10(1), Winter: 69–86, reprinted here in Chapter 13. Kydland, Finn E. and Edward C.Prescott (1997) “A Response [to Milton Friedman],” Journal of Economic Perspectives, 11(1), Winter: 210–11. Lucas, Robert E., Jr. (1972) “Expectations and the Neutrality of Money,” Journal of Economic Theory , 4(2), April: 103–24, reprinted in Lucas (1981): 66–89. Lucas, Robert E., Jr. (1973) “Some Output-Inflation Tradeoffs,” American Economic Review, 63(3), June: 326–34, reprinted in Lucas (1981): 131–45. Lucas, Robert E., Jr. (1975) “An Equilibrium Model of the Business Cycle,” Journal of Political Economy, 83(6): 1113–44, reprinted in Lucas (1981): 179–214. Lucas, Robert E., Jr. (1976) “Econometric Policy Evaluation: A Critique,” in Karl Brunner and Allan H.Meltzer (eds) The Phillips Curve and Labor Markets, vol. 1 of Carnegie-Rochester Conference Series on Public Policy, Amsterdam: North Holland, 19–46, reprinted in Lucas (1981): 104–30. Lucas, Robert E., Jr. (1977) “Understanding Business Cycles,” in Karl Brunner and Allan H.Meltzer (eds) Stabilization of the Domestic and International Economy, Carnegie-Rochester Conference Series in Public Policy, Amsterdam: North Holland, 7–29, reprinted in Lucas (1981): 215–39. Lucas, Robert E., Jr. (1978) “Unemployment Policy,” American Economic Review, 68(2), May: 353–57, reprinted in Lucas (1981): 240–47. Lucas, Robert E., Jr. (1980) “Methods and Problems in Business Cycle Theory,” Journal of Money, Credit and Banking, 12(4), pt. 2, November: 696–713, reprinted in Lucas (1981): 271–96. Lucas, Robert E., Jr. (1981) Studies in Business Cycle Theory, Cambridge, Mass.: MIT Press. Lucas, Robert E., Jr. (1987) Models of Business Cycles, Oxford: Blackwell. Lucas, Robert E., Jr., and Thomas J.Sargent (1979) “After Keynesian Macroeconomics,” Federal Reserve Bank of Minneapolis Quarterly Review, 3(2), reprinted in Robert E.Lucas, Jr. and Thomas J.Sargent (eds) (1981) Rational © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE LIMITS OF BUSINESS CYCLE RESEARCH 41 Expectations and Econometric Practice, Minneapolis: The University of Minnesota Press, 295–319. Mankiw, N.Gregory (1989) “Real Business Cycles: A New Keynesian Perspective,” Journal of Economic Perspectives, 3(3), Summer: 79–90, reprinted here in Chapter 28 . Mantel, R. (1974) “On the Characterization of Aggregate Excess Demand,” Journal of Economic Theory, 7(3): 348–53. Mantel, R. (1976) “Homothetic Preferences and Community Excess Demand Functions,” Journal of Economic Theory, 12(2): 197–201. Manuelli, Rodolfo and Thomas J.Sargent (1988) “Models of Business Cycles: A Review Essay,” Journal of Monetary Economics, 22:523–42. Mas-Collel, A. (1977) “On the Equilibrium Price Set of an Exchange Economy,” Journal of Mathematical Economics, 4(2): 117–26. Nelson, Charles R. and H.Kang (1981) “Spurious Periodicity in Inappropriately Detrended Time Series,” Econometrica, 49(3), May: 741–51. Popper, Karl (1959) The Logic of Scientific Discovery, London: Hutchinson. Popper, Karl (1972) Conjectures and Refutations: The Growth of Scientific Knowledge, 4th ed., London: Routledge. Prescott, Edward C. (1986a) “Theory Ahead of Business Cycle Measurement,” in Federal Reserve Bank of Minneapolis Quarterly Review, 10(4), Fall: 9–22, reprinted here in Chapter 4. Prescott, Edward C. (1986b) “Response to a Skeptic,” Federal Reserve Bank of Minneapolis Quarterly Review, 10(4), Fall: 28–33, reprinted here in Chapter 6. Romer, Christina (1986a) “Is the Stabilization of the Postwar Economy a Figment of the Data?” American Economic Review, 76(3), June: 314–34. Romer, Christina (1986b) “New Estimates of Prewar Gross National Product and Unemployment,” Journal of Economic History, 46:341–52. Romer, Christina (1989) “The Prewar Business Cycle Reconsidered: New Estimates of Gross National Product, 1869–1908,” Journal of Political Economy, 97(1), February: 1–37. Samuelson, Paul A. (1948) Economics: An Introductory Analysis, New York: McGrawHill. Sargent, Thomas J. (1979) Macroeconomic Theory, New York: Academic Press. Sato, R. (1966) “On the Adjustment Time in Neo-Classical Growth Models,” Review of Economic Studies, 33, July: 263–68. Shafer, Wayne and Hugo Sonnenschein. (1982) “Market Demand and Excess Demand Functions,” in K.J.Arrow and M.D.Intriligator (eds) Handbook of Mathematical Economics, vol. 2, Amsterdam: North-Holland: 671–93. Shoven, John B. and John Whalley (1992) Applying General Equilibrium, New York: Cambridge University Press. Siegler, Mark V. (1997) “Real Output and Business Cycle Volatility, 1869– 1993: US Experience in International Perspective,” doctoral dissertation, University of California, Davis. Slutsky, Eugen E. ([1927] 1937) “The Summation of Random Causes as the Source of Cyclic Processes”, Econometrica, 5:105–46. Originally published in Russian in 1927. Smith, Gregor and Stanley Zin (1997) “Real Business Cycle Realizations, 1925– 1995,” Carnegie-Rochester Series in Public Policy, forthcoming. Snowdon, Brian, Howard Vane, and Peter Wynarczyk (eds) (1994) A Modern Guide to Macroeconomics, Aldershot: Edward Elgar. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 42 INTRODUCTION Solow, Robert M. (1956) “A Contribution to the Theory of Economic Growth,” Quarterly Journal of Economics, 70(1), February: 65–94. Solow, Robert M. (1957) “Technical Change and the Aggregate Production Function,” Review of Economics and Statistics, 39(3), August: 312–20, reprinted here in Chapter 27. Solow, Robert M. (1970) Growth Theory: An Exposition, Oxford: Blackwell. Solow, Robert M. (1990) “Reactions to Conference Papers,” in Peter Diamond (ed.) Growth/Productivity/Unemployment: Essays to Celebrate Bob Solow’s Birthday, Cambridge, Mass.: MIT Press: 221–29. Sonnenschein, Hugo (1973) “Do Walras’ Law and Continuity Characterize the Class of Community Excess Demand Functions?” Journal of Economic Theory, 6(4): 345–54. Sonnenschein, Hugo (1974) “Market Excess Demand Functions,” Econometrica, 40(3), May: 549–63. Stoker, Thomas M. (1993) “Empirical Approaches to the Problem of Aggregation over Individuals,” Journal of Economic Literature, 21(4), December: 1827–74. Summers, Lawrence H. (1986) “Some Skeptical Observations on Real Business Cycle Theory,” Federal Reserve Bank of Minneapolis Quarterly Review, 10(4), Fall: 23–27, reprinted here in Chapter 5. Watson, Mark W. (1993) “Measures of Fit for Calibrated Models,” Journal of Political Economy, 101(6), December: 1011–41, reprinted here in Chapter 17. Weir, David (1986) “The Reliability of Historical Macroeconomic Data for Comparing Cyclical Stability,” Journal of Economic History, 46(2), June: 353–65. Yule, G.Udny (1926) “Why Do We Sometimes Get Nonsense-Correlations between Time-Series?” Journal of the Royal Statistical Society, 89(1): 1–65. Reprinted in David F.Hendry and Mary S.Morgan (eds) The Foundations of Econometric Analysis. Cambridge: Cambridge University Press, 1995, ch. 9. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors Chapter 2 A user’s guide to solving real business cycle models The typical real business cycle model is based upon an economy populated by identical infinitely lived households and firms, so that economic choices are reflected in the decisions made by a single representative agent. It is assumed that both output and factor markets are characterized by perfect competition. Households sell capital, kt, to firms at the rental rate of capital, and sell labor, ht, at the real wage rate. Each period, firms choose capital and labor subject to a production function to maximize profits. Output is produced according to a constant-returns-to-scale production function that is subject to random technology shocks. Specifically yt=ztf(kt, ht), where yt is output and zt is the technology shock. (The price of output is normalized to one.) Households’ decisions are more complicated: given their initial capital stock, agents determine how much labor to supply and how much consumption and investment to purchase. These choices are made in order to maximize the expected value of lifetime utility. Households must forecast the future path of wages and the rental rate of capital. It is assumed that these forecasts are made rationally. A rational expectations equilibrium consists of sequences for consumption, capital, labor, output, wages, and the rental rate of capital such that factor and output markets clear. While it is fairly straightforward to show that a competitive equilibrium exists, it is difficult to solve for the equilibrium sequences directly. Instead, an indirect approach is taken in which the Pareto optimum for this economy is determined (this will be unique given the assumption of representative agents). As shown by Debreu (1954), the Pareto optimum as characterized by the optimal sequences for consumption, labor, and capital in this environment will be identical to that in a competitive equilibrium. Furthermore, factor prices are determined by the marginal products of capital and labor evaluated at the equilibrium quantities. (For a detailed exposition of the connection between the competitive equilibrium and Pareto optimum in a real business cycle model, see Prescott, 1986 [4].) We now provide an example of solving such a model. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 44 INTRODUCTION I DERIVING THE EQUILIBRIUM CONDITIONS The first step in solving for the competitive equilibrium is to determine the Pareto optimum. To do this, the real business cycle model is recast as the following social planner’s problem: (2.1) where E1[·] denotes expectations conditional on information at t=1, 0<β <1 is agents’ discount factor, ct denotes consumption, (1-ht) is leisure (agents’ endowment of time is normalized to one), it is investment, and 0<δ<1 is the depreciation rate of capital. The exogenous technology shock is assumed to follow the autoregressive process given in the last equation; the autocorrelation parameter is 0ⱕρⱕ1 and the innovation to technology is assumed to have a mean of one and standard deviation σε. The first two constraints in equation (2.1) are the economy-wide resource constraint, and the second is the law of motion for the capital stock. Dynamic programming problem This infinite horizon problem can be solved by exploiting its recursive structure. That is, the nature of the social planner’s problem is the same every period: given the beginning-of-period capital stock and the current technology shock, choose consumption, labor, and investment. Note that utility is assumed to be time-separable: that is, the choices of consumption and labor at time t do not affect the marginal utilities of consumption and leisure in any other time period. Because of this recursive structure, it is (2.2) © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors GUIDE TO SOLVING REAL BUSINESS CYCLES 45 useful to cast the maximization problem as the following dynamic programming problem (for a discussion of dynamic programming, see Sargent (1987)): (Note that investment has been eliminated by using the law of motion for the capital stock.) A solution to this problem must satisfy the following necessary conditions and resource constraint: Where the notation Ui,t; i=l, 2 denotes the derivative of the utility function with respect to the ith argument evaluated at the quantities (ct, 1-ht); fi,t; i=1,2 has an analogous interpretation. N1 represents the intratemporal efficiency condition (the labor-leisure tradeoff). It implies that the marginal rate of substitution between labor and consumption must equal the marginal product of labor. The second condition, N2, represents the intertemporal efficiency condition. The left-hand side represents the marginal cost in terms of utility of investing in more capital, while the right-hand side represents the expected marginal utility gain; at an optimum, these costs and benefits must be equal. To simplify the analysis (again, see Prescott (1986 [4]) for a justification), assume the following functional forms: (The assumption that utility is linear in leisure is based on Hansen’s (1985 [8]) model.) Then the three equilibrium conditions become: (2.3) A steady-state equilibrium for this economy is one in which the technology shock is assumed to be constant, so that there is no uncertainty: that is, zt=1 for all t, and the values of capital, labor, and consumption are for all t. Imposing these steady-state constant, conditions in equation (2.3), the steady-state values are found by solving the following steady-state equilibrium conditions: In the above expressions, denotes the steady-state level of output. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 46 INTRODUCTION Calibration The next step in solving the model is to choose parameter values for the model. This is done through calibration: the set of parameters (δ, ß, A, α) are chosen so that the steady-state behavior of the model matches the long-run characteristics of the data. The features of the data which do not exhibit cyclical characteristics are: 1 2 3 4 (1-α)=labor’s average share of output. ß-1-1=average risk-free real interest rate. Given (α, ß) choose d so that the output-capital ratio (from (SS2) is consistent with observation. The parameter A determines the time spent in work activity. To see this, multiply both sides of (SS1) by and rearrange the expression to But the steady-state resource constraint, yield: so that the output-consump(SS3), implies that tion ratio is implied by the parameter values chosen in the previous three steps. Hence, the choice of A directly determines Typical parameter values based on postwar US data (see Hansen and Wright (1992 [9])) are: α=0.36 implying labor’s share is 64 percent, ß=0.99 implying an annual riskless interest rate of 0.04 percent, δ=0.025 implying the capital-output ratio (where output is measured on a quarterly basis) of roughly 10, and A=3 which implies that roughly 30 percent of time is spent in work activity. (These values will be used in Section IV below.) II LINEARIZATION The solution to the social planner’s problem is characterized by a set of policy functions for capital, consumption, and labor; moreover, the solution exists and is unique (see Prescott (1986 [4])). There is, however, no analytical solution. To make the model operational, therefore, an approximate numerical solution is found. One of the simplest methods is to take a linear approximation (i.e., a first-order Taylor series expansion) of the three equilibrium conditions and the law of motion of the technology Provided the stochastic shock around the steady-state values behavior of the model does not push the economy too far from the steadystate behavior, the linear approximation will be a good one. (The discussion below follows closely that of Farmer (1993).) This technique is demonstrated below.1 Intratemporal efficiency condition The optimal labor-leisure choice is represented by condition N1: Linearizing around the steady-state values : © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors GUIDE TO SOLVING REAL BUSINESS CYCLES 47 Note that in the last expression, all variables have been expressed as percentage deviations from the steady state (the first two terms modify the in steady respective derivatives while the last term uses the fact that state). Consumption can be expressed as a percentage deviation from steady condition dividing state by using the steady-state both sides of the equation by this expression and denoting percentage deviations from steady state as equation (2.4) can be written as: (2.5) Intertemporal efficiency condition This efficiency condition is given by N2: Again, linearizing around the steady state, and expressing all variables as percentage deviations from steady state, yields: Multiplying each side of the equation by and using the steady-state condition (SS2) that (2.6) © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 48 INTRODUCTION Resource constraint Following the same procedure as before, linearizing the resource constraint around the steady state yields: (2.7) Technology shock process The critical difference between the steady-state model and the real business cycle model is the assumption that technology shocks are random—the shocks follow the autoregressive process described in equation (2.1). Linearizing the autoregressive process for the technology shock results in: (2.8) Taking expectations of both sides: (2.9) III SOLUTION METHOD The equations that define a rational expectations equilibrium (equations (2.5), (2.6), (2.7), and (2.9)) can be written as a vector expectational difference equation. Let where bold print denotes a vector, then the linear system of equations can be written as: (2.10) The matrices A and B are: © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors GUIDE TO SOLVING REAL BUSINESS CYCLES 49 Premultiplying both sides of equation (2.10) by A-1 yields: (2.11) The matrix A-1B can be decomposed as (see Hamilton (1994) for details): (2.12) where Q is a matrix whose columns are the eigenvectors of A-1B and ⌳ is a diagonal matrix whose diagonal elements are the eigenvalues of A-1B. Using this decomposition and premultiplying both sides of the resulting expression in equation (2.11) by Q-1 yields: (2.13) Note that the elements of the defined (4×1) column vector dt are constructed from a linear combination of the elements in the rows of the (4× 4) matrix Q-1 and the elements of the (4×1) column vector ut. Since ⌳ is a diagonal matrix, equation (2.13) implies four independent equations: (2.14) Since the equations in equation (2.14) must hold every period, it is possible to recursively substitute the expressions forward for T periods to yield: (2.15) The λi are four distinct eigenvalues associated with the four equilibrium conditions (equations (2.5)–(2.8)). Since one of these conditions is the law of motion for the exogenous technology shock (equation (2.8)), one of the eigenvalues will be ρ-1. Also, the first rows of the matrices A and B are determined by the intratemporal efficiency condition; since this is not a dynamic relationship, one of the eigenvalues will be zero. The remaining two eigenvalues will bracket the value of unity as is typical for a saddle path equilibrium implied by the underlying stochastic growth framework. As implied by equation (2.15), the stable, rational expectations solution to the expectational difference equation is associated with the eigenvalue with a value less than one. That is, if λi>1, then iterating forward implies di,t→∞ which is not a permissible equilibrium. Furthermore, for equation (2.15) to hold for all T (again taking the limit of the right-hand side), in the stable case when, λ<1, it must be the true that di,t=0; this restriction provides the desired solution. That is, di,t=0 imposes the linear restriction which is consistent with a rational expectations solution. on (Recall that di,t represents a linear combination between the elements of a particular row of Q-1 and the elements of the vector ut.) IV A PARAMETRIC EXAMPLE In this section, a parameterized version of the RBC model described above is solved. The following parameter values are used: (ß=0.99, © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 50 INTRODUCTION α=0.36, δ=0.025, A=3). These imply the following steady-state values: Note that these values imply that agents spend roughly 30 percent of their time in work activities and the capital-output ratio is approximately 10 (output is measured on a quarterly basis); both of these values are broadly consistent with US experience (see McGrattan, 1994). The remaining parameter values determine the behavior of the technology shock. These are estimated by constructing the Solow residual2 and then detrending that series linearly. Specifically, the Solow residual is defined as Zt=1n yt-α 1n kt-(1-α) 1n ht. The Zt series can then be regressed on a linear time trend (which is consistent with the assumption of constant technological progress) and the residual is identified as the technology shock zt. Using this procedure on quarterly data over the period 60.1-94.4 resulted in an estimate of the serial correlation of zt (the parameter ?) to be 0.95. The variance of the shock to technology (i.e., the variance of et in equation (2.8)) was estimated to be 0.007. Note that the variance of the technology shock is not relevant in solving the linearized version of the model—however, when the solution of the model is used to generate artificial time series in the simulation of the economy, this parameter value must be stipulated. These values generated the following entries into the A and B matrices: Following the steps described in the previous section (premultiplying by A-1) yields the following: Next, decomposing A-1 B into Q⌳ ⌳ Q-1 and then premultiplying by Q-1 yields © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors GUIDE TO SOLVING REAL BUSINESS CYCLES 51 The entries in the matrix ⌳ (i.e., the eigenvalues of A-1B) determine the solution. Note that the second diagonal entry is (accounting for rounding error) ρ-1. The fourth row of ⌳ is associated with the intratemporal efficiency condition. These values are proportional to those given in the first row of the A matrix; consequently, dividing all entries by (-2.62) returns the original intratemporal efficiency condition. The remaining two entries in the ⌳ matrix are those related to the saddle path properties of the steady-state solution. Since a stable rational expectations solution is associated with an eigenvalue less than unity, the third row of the Q-1 matrix provides the linear restriction we are seeking. That is, the rational expectations solution is: (2.16) The law of motion for the capital stock (the parameter values are given in the third row of the A matrix) and the intratemporal efficiency condition provides two more equilibrium conditions: (2.17) (2.18) A random number generator can next be used to produce a sequence of technology shocks. The above equilibrium equations can then be used to produce time series for capital, consumption, labor, and output. V ANALYZING OUTPUT FROM THE ARTIFICIAL ECONOMY The solution to the model is characterized by equations (2.16)–(2.18). Given initial values for capital, and next generating a path for the exogenous these equations will produce time series for technology shock Two other series that most macroeconomists are interested in, namely, output and investment, can be generated by linearizing the production function and the resource constraint, respectively. Specifically, for output, linearizing the assumed Cobb-Douglas and using the calibrated value production function (i.e., that α= 0.36) yields the following equation: (2.19) Finally, a linear approximation of the condition that, in equilibrium, output must equal the sum of consumption and investment, can be expressed in the form of a percentage deviation from the steady state as: © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 52 INTRODUCTION (2.20) Using the steady-state values employed in the numerical solution, the investment equation becomes: (2.21) Hence, equilibrium in this economy is described by the following set of equations: To generate the time series implied by the model, it is necessary first to generate a series for the innovations to the technology shock, i.e., These Figure 2.1 Output, Consumption, and Investment in RBC Model © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors GUIDE TO SOLVING REAL BUSINESS CYCLES 53 Table 2.1 Descriptive Statistics for US and RBC model Source: Statistics for US data are taken from Kydland and Prescott (1990 [20]), tables I and II, pp. 10–11. are assumed to have a mean of zero and a variance that is consistent with the observed variance for the innovations, which, as mentioned above, is and using a random number roughly 0.007. Then, initializing generator in order to generate the innovations, a path for the technology shocks is created. Next, assuming that all remaining values are initially at their steady state (which implies that all initial values are set to zero), the system of equations above can be solved to produce the time path for the endogenous variables. We generate artificial time paths for consumption, output, and investment (3000 observations were created and only the last 120 were examined), and these are shown in Figure 2.1. It is clear from Figure 2.1, as is also true in the actual data, that the volatility of investment is greater than that of output, which is greater than that of consumption. To see this more precisely, the standard deviation of consumption, labor, and investment relative to output is reported in Table 2.1, along with the correlations of these series with output. NOTES 1 2 Recall that the general form for the Taylor series expansion of a function around a point x* is: where N! denotes factorial. The use of the Solow residual as a measure of technology shocks is discussed in Hoover and Salyer (1996). REFERENCES Debreu, Gerard (1954) “Valuation Equilibrium and Pareto Optimum,” Proceedings of the National Academy of Science, 40:588–92. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 54 INTRODUCTION Farmer, Roger E.A. (1993) The Macroeconomics of Self-fulfilling Prophecies, Cambridge, Mass.: MIT Press. Hamilton, James D. (1994) Time Series Analysis, Princeton: Princeton University Press. Hansen, Gary D. (1985) “Indivisible Labor and the Business Cycle,” Journal of Monetary Economics, 16(3), November: 309–28, reprinted here in Chapter 8. Hansen, Gary D. and Randall Wright (1992) “The Labor Market in Real Business Cycle Theory,” Federal Reserve Bank of Minneapolis Quarterly Review, 16(2), Spring: 2–12, reprinted here in Chapter 9. Hoover, Kevin D. and Kevin D.Salyer (1996) “Technology Shocks or Colored Noise? Why Real Business Cycle Models Cannot Explain Actual Business Cycles,” unpublished manuscript. Kydland, Finn E. and Edward C.Prescott (1990) “Business Cycles: Real Facts and a Monetary Myth,” Federal Reserve Bank of Minneapolis Quarterly Review, 14(2), Spring: 3–18, reprinted here in Chapter 20. McGratten, Ellen R. (1994) “A Progress Report on Business Cycle Models,” Federal Reserve Bank of Minneapolis Quarterly Review, 18(4), Fall: 2–16. Prescott, Edward C. (1986) “Theory Ahead of Business Cycle Measurement,” Federal Reserve Bank of Minneapolis Quarterly Review, 10(4), Fall: 9– 22, reprinted here in Chapter 4. Sargent, Thomas J. (1987) Dynamic Macroeconomic Theory, Cambridge, Mass.: Harvard University Press. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors Part II The foundations of real business cycle modeling © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors CHAPTER 3 57 ECONOMETRICA VOLUME 50 NOVEMBER, 1982 NUMBER 6 TIME TO BUILD AND AGGREGATE FLUCTUATIONS BY FINN E.KYDLAND AND EDWARD C.PRESCOTT1 The equilibrium growth model is modified and used to explain the cyclical variances of a set of economic time series, the covariances between real output and the other series, and the autocovariance of output. The model is fitted to quarterly data for the post-war U.S. economy. Crucial features of the model are the assumption that more than one time period is required for the construction of new productive capital, and the non-time-separable utility function that admits greater intertemporal substitution of leisure. The fit is surprisingly good in light of the model’s simplicity and the small number of free parameters. 1. INTRODUCTION THAT WINE IS NOT MADE in a day has long been recognized by economists (e.g., Böhm–Bawerk [6]). But, neither are ships nor factories built in a day. A thesis of this essay is that the assumption of multiple-period construction is crucial for explaining aggregate fluctuations. A general equilibrium model is developed and fitted to U.S. quarterly data for the post-war period. The co-movements of the fluctuations for the fitted model are quantitatively consistent with the corresponding co-movements for U.S. data. In addition, the serial correlations of cyclical output for the model match well with those observed. Our approach integrates growth and business cycle theory. Like standard growth theory, a representative infinitely-lived household is assumed. As fluctuations in employment are central to the business cycle, the stand-in consumer values not only consumption but also leisure. One very important modification to the standard growth model is that multiple periods are required to build new capital goods and only finished capital goods are part of the productive capital stock. Each stage of production requires a period and utilizes resources. Half-finished ships and factories are not part of the productive capital stock. Section 2 contains a short critique of the commonly used investment technologies, and presents evidence that single-period production, even with adjustment costs, is inadequate. The preference-technologyinformation structure of the model is presented in Section 3. A crucial feature of preferences is the non-time-separable utility function that admits greater intertemporal substitution of leisure. The exogenous stochastic components in the model are shocks to technology and imperfect indicators of productivity. The two technology shocks differ in their persistence. The steady state for the model is determined in Section 4, and quadratic approximations are made which result in an “indirect” quadratic utility function that 1 The research was supported by the National Science Foundation. We are grateful to Scan Becketti, Fischer Black, Robert S.Chirinko, Mark Gersovitz, Christopher A.Sims, and John B. Taylor for helpful comments, to Sumru Altug for research assistance, and to the participants in the seminars at the several universities at which earlier drafts were presented. 1345 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 58 FOUNDATIONS 1346 F.E.KYDLAND AND E.C.PRESCOTT values leisure, the capital goods, and the negative of investments. Most of the relatively small number of parameters are estimated using steady state considerations. Findings in other applied areas of economics are also used to calibrate the model. For example, the assumed number of periods required to build new productive capital is of the magnitude reported by business, and findings in labor economics are used to restrict the utility function. The small set of free parameters imposes considerable discipline upon the inquiry. The estimated model and the comparison of its predictions with the empirical regularities of interest are in Section 5. The final section contains concluding comments. 2. A CRITIQUE OF CONVENTIONAL AGGREGATE INVESTMENT TECHNOLOGIES There are two basic technologies that have been adopted in empirical studies of aggregate investment behavior. The first assumes a constant-returns-to-scale neoclassical production function F with labor L and capital K as the inputs. Total output F(K, L) constrains the sum of investment and consumption, or C+I≤F(K, L), where C, I, K, L≥0. The rate of change of capital, is investment less depreciation, and depreciation is proportional with factor δ to the capital stock, that is, This is the technology underlying the work of Jorgenson [19] on investment behavior. An implication of this technology is that the relative price of the investment and consumption goods will be a constant independent of the relative outputs of the two goods.2 It also implies that the shadow price of existing capital will be the same as the price of the investment good.3 There is a sizable empirical literature that has found a strong association between the level of investment and a shadow price of capital obtained from stock market data (see [26]). This finding is inconsistent with this assumed technology as is the fact that this shadow price varies considerably over the business cycle. The alternative technology, which is consistent with these findings, is the single capital good adjustment cost technology.4 Much of that literature is based upon the problem facing the firm and the aggregation problem receives little attention. This has led some to distinguish between internal and external adjustment costs. For aggregate investment theory this is not an issue (see [29]) though for other questions it will be. Labor resources are needed to install capital whether the acquiring or supplying firm installs the equipment. With competitive equilibrium it is the aggregate production 2 This, of course, assumes neither C nor I is zero. Sargent [32], within a growth context with shocks to both preferences and technology, has at a theoretical level analyzed the equilibrium with corners. Only when investment was zero did the price of the investment good relative to that of the consumption good become different from one and then it was less than one. This was not an empirical study and Sargent states that there currently are no computationally practical econometric methods for conducting an empirical investigation within that theoretical framework. 3 The shadow price of capital has been emphasized by Brunner and Meltzer [7] and Tobin [36] in their aggregate models. 4 See [1, 17] for recent empirical studies based on this technology. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors TIME TO BUILD AGGREGATE FLUCTUATIONS 59 1347 possibility set that matters. That is, if the Yj are the production possibility sets of the firms associated with a given industrial organization and for some other industrial organization, the same aggregate supply behavior results if The adjustment cost model, rather than assuming a linear product transformation curve between the investment and consumption goods, imposes curvature. This can be represented by the following technology: where G like F is increasing, concave, and homogeneous of degree one. Letting the price of the consumption good be one, the price of the investment good qt, the rental price of capital rt, and the wage rate wt, the firm’s problem is to maximize real profits, Ct+qtIt-wtLt-rtKt, subject to the production constraint. As constant returns to scale are assumed, the distribution of capital does not matter, and one can proceed as if there were a single price-taking firm. Assuming an interior solution, given that this technology displays constant returns to scale and that the technology is separable between inputs and outputs, it follows that It=F(Kt, Lt)h(qt)=Zth(qt), where Zt is defined to be aggregate output. The function h is increasing, so high investment-output ratios are associated with a high price of the investment good relative to the consumption good. Figure 1 depicts the investment-consumption product transformation curve and Figure 2 the function h(q). For any I/Z, the negative of the slope of the transformation curve in Figure 1 is the height of the curve in Figure 2. This establishes that a higher q will be associated with higher investment for this technology. This restriction of the theory is consistent with the empirical findings previously cited. There are other predictions of this theory, however, which are questionable. If we think of the q-investment curve h depicted in Figure 2 as a supply curve, the short- and the long-run supply elasticities will be equal. Typically, economists argue that there are specialized resources which cannot be instantaneously and costlessly transferred between industries and that even though short-run elasticities may be low, in the long run supply elasticities are high. As there are no specialized resources for the adjustment cost technology, such considerations are absent and there are no penalties resulting from rapid adjustment in the relative outputs of the consumption and investment good. FIGURE 1. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors FIGURE 2. 60 FOUNDATIONS 1348 F.E.KYDLAND AND E.C.PRESCOTT To test whether the theory is a reasonable approximation, we examined crosssection state data. The correlations between the ratios of commercial construction to either state personal income or state employment and price per square foot5 are both -0.35. With perfectly elastic supply and uncorrelated supply and demand errors, this correlation cannot be positive. To explain this large negative correlation, one needs a combination of high variability in the cross-sectional supply relative to cross-sectional demand plus a positive slope for the supply curve. Our view is that, given mobility of resources, it seems more plausible that the demand is the more variable. Admitting potential data problems, this cross-sectional result casts some doubt upon the adequacy of the single capital good adjustment cost model. At the aggregate level, an implication of the single capital good adjustment cost model is that when the investment-output ratio is regressed on current and lagged q, only current q should matter.6 The findings in [26] are counter to this prediction. In summary, our view is that neither the neoclassical nor the adjustment cost technologies are adequate. The neoclassical structure is inconsistent with the positive association between the shadow price of capital and investment activity. The adjustment cost technology is consistent with this observation, but inconsistent with cross-sectional data and the association of investment with the lagged as well as the current capital shadow prices. In addition, the implication that longand short-run supply elasticities are equal is one which we think a technology should not have. Most destructive of all to the adjustment-cost technology, however, is the finding that the time required to complete investment projects is not short relative to the business cycle. Mayer [27], on the basis of a survey, found that the average time (weighted by the size of the project) between the decision to undertake an investment project and the completion of it was twenty-one months. Similarly, Hall [13] found the average lag between the design of a project and when it becomes productive to be about two years. It is a thesis of this essay that periods this long or even half that long have important effects upon the serial correlation properties of the cyclical components of investment and total output as well as on certain co-movements of aggregate variables. The technological requirement that there are multiple stages of production is not the delivery lag problem considered by Jorgenson [19]. He theorized at the firm level and imposed no consistency of behavior requirement for suppliers and demanders of the investment good. His was not a market equilibrium analysis and there was no theory accounting for the delivery lag. Developing such a market theory with information asymmetries, queues, rationing, and the like is a challenging problem confronting students of industrial organization. 5 The data on commercial construction and price per square foot were for 1978 and were obtained from F.W.Dodge Division of McGraw-Hill. 6 This observation is due to Fumio Hayashi. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors TIME TO BUILD AGGREGATE FLUCTUATIONS 61 1349 Our technology assumes that a single period is required for each stage of construction or that the time required to build new capital is a constant. This is not to argue that there are not alternative technologies with different construction periods, patterns of resource use, and total costs. We have found no evidence that the capital goods are built significantly more rapidly when total investment activity is higher or lower. Lengthening delivery lags (see [9]) in periods of high activity may be a matter of longer queues and actual construction times may be shorter. Premiums paid for earlier delivery could very well be for a more advanced position in the queue than for a more rapidly constructed factory. These are, of course, empirical questions, and important cyclical variation in the construction period would necessitate an alternative technology. Our time-to-build technology is consistent with short-run fluctuations in the shadow price of capital because in the short run capital is supplied inelastically. It also implies that the long-run supply is infinitely elastic, so on average the relative price of the investment good is independent of the investment-output ratio. 3. THE MODEL Technology The technology assumes time is required to build new productive capital. Let Sjt be the number of projects j stages or j periods from completion for j =1, …, J-1, where J periods are required to build new productive capacity. New investment projects initiated in period t are SJt. The recursive representation of the laws of motion of these capital stocks is (3.1) (3.2) Here, kt is the capital stock at the beginning of period t, and δ is the depreciation rate. The element SJt is a decision variable for period t. The final capital good is the inventory stock yt inherited from the previous period.7 Thus, in this economy, there are J+1 types of capital: inventories yt, productive capital kt, and the capital stocks j stages from completion for j=1, …, J-1. These variables summarize the effects of past decisions upon current and future production possibilities. Let j for j=1, …, J be the fraction of the resources allocated to the investment project in the jth stage from the last. Total non-inventory investment in period t is Total investment, i t, is this amount plus inventory 7 All stocks are beginning-of-the-period stocks. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 62 FOUNDATIONS 1350 F.E.KYDLAND AND E.C.PRESCOTT investment yt+1-yt, and consequently (3.3) Total output, that is, the sum of consumption ct and investment, is constrained as follows: (3.4) where nt is labor input, t a shock to technology, and f is a constant-returns-to-scale production function to be parameterized subsequently. Treating inventories as a factor of production warrants some discussion. With larger inventories, stores can economize on labor resources allocated to restocking. Firms, by making larger production runs, reduce equipment down time associated with shifting from producing one type of good to another. Besides considerations such as these, analytic considerations necessitated this approach. If inventories were not a factor of production, it would be impossible to locally approximate the economy using a quadratic objective and linear constraints. Without such an approximation no practical computational method currently exists for computing the equilibrium process of the model. The production function is assumed to have the form (3.5) where 0<<1, 0<σ<1, and 0<<∞. This form was selected because, among other things, it results in a share for labor in the steady state. The elasticity of substitution between capital and inventory is 1/(1+ν). This elasticity is probably less than one which is why ν is required to be positive. Preferences The preference function, whose expected value the representative household maximizes, has the form where 0<β<1 is the discount factor, lt leisure, L the lag operator, and Normalizing so that one is the endowment of time, we let nt=1-lt be the time allocated to market activity. The polynomial lag operator is restricted so that the αi sum to one, and αi=(1-η)i-1α1 for i≥1, where 0<η≤1. With these restrictions, By defining the variable recursive representation: the distributed lag has the following (3.6) © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors TIME TO BUILD AGGREGATE FLUCTUATIONS 63 1351 The variable at summarizes the effects of all past leisure choices on current and future preferences. If ns=nt for all s≤t, then at=nt/η, and the distributed lag is simply 1-nt. The parameters α0 and determine the degree to which leisure is intertemporally substitutable. We require 0<η≤1 and 0<α0≤1. The nearer a0 is to one, the less is the intertemporal substitution of leisure. For α0 equal to one, time-separable utility results. With η equal to one, at equals nt-1. This is the structure employed in [33]. As η approaches zero, past leisure choices have greater effect upon current utility flows. Non-time-separable utility functions are implicit in the empirical study of aggregate labor supply in [25]. Grossman [12] and Lucas [24] discuss why a non-time-separable utility function is needed to explain the business cycle fluctuations in employment and consumption. A micro justification for our hypothesized structure based on a Beckerian household production function is as follows.8 Time allocated to non-market activities, that is lt, is used in household production. If there is a stock of household projects with varying output per unit of time, the rational household would allocate lt to those projects with the greatest returns per time unit. If the household has allocated a larger amount of time to non-market activities in the recent past, then only projects with smaller yields should remain. Thus, if at is lower, the marginal utility value of lt should be smaller. Cross-sectional evidence of households’ willingness to redistribute labor supply over time is the lumpiness of that supply. There are vacations and movements of household members into and out of the labor force for extended periods which are not in response to large movements in the real wage. Another observation suggesting high intertemporal substitutability of leisure is the large seasonal variation in hours of market employment. Finally, the failure of Abowd and Ashenfelter [2] to find a significant wage premium for jobs with more variable employment and earnings patterns is further evidence. In summary, household production theory and crosssectional evidence support a non-time-separable utility function that admits greater intertemporal substitution of leisure—something which is needed to explain aggregate movements in employment in an equilibrium model. The utility function in our model is assumed to have the form where γ<1 and γ0. If the term in the square brackets is interpreted as a composite commodity, then this is the constant-relative-risk-aversion utility function with the relative degree of risk aversion being 1-γ. We thought this composite commodity should be homogeneous of degree one as is the case when there is a single good. The relative size of the two exponents inside the brackets is 8 We thank Nasser Saïdi for suggesting this argument. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 64 FOUNDATIONS 1352 F.E.KYDLAND AND E.C.PRESCOTT motivated by the fact that households’ allocation of time to nonmarket activities is about twice as large as the allocation to market activities. Information Structure We assume that the technology parameter is subject to a stochastic process with components of differing persistence. The productivity parameter is not observed but the stand-in consumer does observe an indicator or noisy measure of this parameter at the beginning of the period. This might be due to errors in reporting data or just the fact that there are errors in the best or consensus forecast of what productivity will be for the period. On the basis of the indicator and knowledge of the economy-wide state variables, decisions of how many new investment projects to initiate and of how much of the time endowment to allocate to the production of marketed goods are made. Subsequent to observing aggregate output, the consumption level is chosen with inventory investment being aggregate output less fixed investment and consumption. Specifically, the technology shock, λt, is the sum of a permanent component, λ1t, and a transitory component,9 λ2t: (3.7) In the spirit of the Friedman-Muth permanent-income model, the permanent component is highly persistent so (3.8) where ρ is less than but near one and ζ1t is a permanent shock.10 The transitory component equals the transitory shock so (3.9) The indicator of productivity, πt, is the sum of actual productivity λt and a third shock ζ3t: (3.10) The shock vectors ζt=(ζ1t, ζ2t, ζ3t) are independent multivariate normal with mean vector zero and diagonal covariance matrix. The period-t labor supply decision nt and new investment project decision sJt are made contingent upon the past history of productivity shocks, the λk for k<t, the indicator of productivity πt, the stocks of capital inherited from the past, and variable 9 The importance of permanent and transitory shocks in studying macro fluctuations is emphasized in [8]. 10 The value used for ρ in this study was 0.95. The reason we restricted ρ to be strictly less than one was technical. The theorem we employ to guarantee the existence of competitive equilibrium requires stationarity of the shock. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors TIME TO BUILD AGGREGATE FLUCTUATIONS 65 1353 at. These decisions cannot be contingent upon λt for it is not observed or deducible at the time of these decisions. The consumption-inventory investment decision, however, is contingent upon λt for aggregate output is observed prior to this decision and λt can be deduced from aggregate output and knowledge of inputs. The state space is an appropriate formalism for representing this recursive information structure. Because of the two-stage decision process, it is not a direct application of Kalman filtering. Like that approach the separation of estimation and control is exploited. The general structure assumes an unobservable state vector, say xt, that follows a vector autoregressive process with independent multivariate normal innovations: (3.11) Observed prior to selecting the first set of decisions is (3.12) The element B1 is a matrix and the e1t are independent over time. Observed prior to the second set of decisions and subsequent to the first set is (3.13) Equations (3.11)–(3.13) define the general information structure. To map our information structure into the general formulation, let B1=[1 1], B2=[1 1], V1=[var(3)], and V2=[0]. With these definitions, the information structure (3.7)– (3.10) viewed as deviations from the mean and the representation (3.11)– (3.13) are equivalent. Let m0t be the expected value and 0 the covariance of the distribution of xt conditional upon the pk=(p1k, p2k) for k<t. Using the conditional probability laws for the multivariate normal distribution (see [28, p. 208]) and letting m1t and 1 be the mean and covariance of xt conditional upon p1t as well, we obtain (3.14) (3.15) Similarly, the mean vector m2t and covariance matrix 2 conditional upon p2t as well are (3.16) (3.17) © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 66 FOUNDATIONS 1354 F.E.KYDLAND AND E.C.PRESCOTT Finally, from (3.11), (3.18) (3.19) The covariances 0, 1, and 2 are defined recursively by (3.15), (3.17), and (3.19). The matrix V0 being of full rank along with the stability of A are sufficient to insure that the method of successive approximations converges exponentially fast to a unique solution. The covariance elements 0, 1, and 2 do not change over time and are therefore not part of the information set. The m0t, m1t, and m2t do change but are sufficient relative to the relevant histories for forecasting future values of both the unobserved state and the observable pt, >t, and for estimating the current unobserved state. Equilibrium To determine the equilibrium process for this model, we exploit the well-known result that, in the absence of externalities, competitive equilibria are Pareto optima. With homogeneous individuals, the relevant Pareto optimum is the one which maximizes the welfare of the stand-in consumer subject to the technology constraints and the information structure. Thus, the problem is to subject to constraints (3.1)–(3.4), (3.6), and (3.11)–(3.13), given k0, s10, …, sJ-1,0, a0, and that x0~N(m0, 0). The decision variables at time t are nt, sJt, ct, and yt+1. Further, nt and sJt cannot be contingent upon p2t for it is observed subsequent to these decisions. This is a standard discounted dynamic programming problem. There are optimal time-invariant or stationary rules of the form nt=n(kt, s1t, s2t, …, sJ-1,t, yt, at, m1t), sJt=s(kt, s1t, s2t, …, SJ-1,t, yt, at, m1t), ct=c(kt, s1t, s2t, …, sJt, yt, at, nt, m2t), yt+1=y(kt, s1t, s2t, …, sJt, yt, at, nt, m2t). It is important to note that the second pair of decisions are contingent upon m2t rather than m1t and that they are contingent also upon the first set of decisions sJt and nt. The existence of such decision rules and the connection with the competitive allocation is established in [31]. But, approximations are necessary before equilibrium decision rules can be computed. Our approach is to determine the steady © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors TIME TO BUILD AGGREGATE FLUCTUATIONS 67 1355 state for the model with no shocks to technology. Next, quadratic approximations are made in the neighborhood of the steady state. Equilibrium decision rules for the resulting approximate economy are then computed. These rules are linear, so in equilibrium the approximate economy is generated by a system of stochastic difference equations for which covariances are easily determined. 4. STEADY STATE, APPROXIMATION, AND COMPUTATION OF EQUILIBRIUM Variables without subscript denote steady state values. The steady state interest rate is r=(1- )/ , and the steady state price of (non-inventory) capital The latter is obtained by observing that 1 units of consumption must be foregone in the current period, 2 units the period before, etc., in order to obtain one additional unit of capital for use next period. Two steady state conditions are obtained by equating marginal products to rental rates, namely fy=r and fk=q(r+). These imply fk/fy=q(r+)/r. For production function (3.5), this reduces to (4.1) Differentiating the production function with respect to capital, substituting for y from (4.1), and equating to the steady-state rental price, one obtains where Solving for k as a function of n yields (4.2) Steady-state output as a function of n is steady state, net investment is zero, so In the (4.3) The steady-state values of c, k, and y are all proportional to We also note that the capital-output ratio is b3/b4, and that consumption’s share to total steady-state output is 1-(b3/b4). Turning now to the consumer’s problem and letting µ be the Lagrange multiplier for the budget constraint and wt the real wage, first-order conditions are © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 68 FOUNDATIONS 1356 F.E.KYDLAND AND E.C.PRESCOTT In the steady state, ct=c, lt=l, and wt=w for all t. Making these substitutions and using the fact that the i sum to one, these expressions simplify to Eliminating µ from these equations yields and l=1-n, this in turn implies Since (4.4) Returning to the production side, the marginal product of labor equals the real wage: (4.5) Using (4.3) and (4.5), we can solve (4.4) for n: That n does not depend upon average matches well with the American experience over the last thirty years. During this period, output per man-hour has increased by a few hundred per cent, yet man-hours per person in the 16–65 age group has changed but a few per cent. Approximation About the Steady State If the utility function u were quadratic and the production function f linear, there would be no need for approximations. In equilibrium, consumption must be equal to output minus investment. We exploit this fact to eliminate the nonlinearity in the constraint set by substituting f(, k, n, y)-i for c in the utility function to obtain u(f(, k, n, y)-i, n, a). The next step is to approximate this function by a quadratic in the neighborhood of the model’s steady state. As investment i is linear in the decision and state variables, it can be eliminated subsequent to the approximation and still preserve a quadratic objective. Consider the general problem of approximating function u(x) near x-. The approximate quadratic function is where x, b ∈ Rn and Q is an n×n symmetric matrix. We want an approximation that is good not only at x- but also at other x in the range experienced during the sample period. Let zi be a vector, all of whose components are zero except for Our approach is to select the elements bi and qii so that the approximation © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors TIME TO BUILD AGGREGATE FLUCTUATIONS 69 1357 error is zero at the and where the selected correspond to the approximate average deviations of the xi from their steady state values The used for , k, y, n, i, and a were 3, 1, 2, 3, 8, and 0.5 per cent, values of respectively.11 The approximation errors being zero at the and requires that The elements qij, i ≠j, are selected to minimize the sum of the squared approximation errors at and The approximation error at the first point is Summing over the square of this error and the three others, differentiating with respect to qij, setting the resulting expression equal to zero and solving for qij, we obtain for i ≠j. Computation of Equilibrium The equilibrium process for the approximate economy maximizes the welfare of the representative household subject to the technological and informational constraints as there are no externalities. This simplifies the determination of the equilibrium process by reducing it to solving a linear-quadratic maximization problem. For such mathematical structure there is a separation of estimation and control. Consequently, the first step in determining the equilibrium decision rules for the approximate economy is to solve the following deterministic problem: 11 We experimented a little and found that the results were essentially the same when the second order Taylor series approximation was used rather than this function. Larry Christiano [10] has found that the quadratic approximation method that we employed yields approximate solutions that are very accurate, even with large variability, for a structure that, like ours, is of the constant elasticity variety. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 70 FOUNDATIONS 1358 F.E.KYDLAND AND E.C.PRESCOTT subject to (4.6) (4.7) (4.8) (4.9) (4.10) (4.11) At this stage, the fact that there is an additive stochastic term in the equation determining xt+1 is ignored as is the fact that xt is not observed for our economy. Constraints (4.6)–(4.9) are the laws of motion for the state variables. The free decision variables are nt, sJt, and yt+1. It was convenient to use inventories taken into the subsequent period, yt+1, as a period t decision variable rather than it because the decisions on inventory carry-over and consumption are made subsequent to the labor supply and new project decisions nt and SJt. For notational simplicity we let the set of state variables other than the unobserved xt be z t=(kt, yt, at , s1t , . . . , sJ-1,t ) and the set of decision variables d t =(n t , s Jt , y t+1 ). The unobserved state variables xt=(x1t , x2t) are the permanent and transitory shocks to technology. Finally, v(x, z) is the value of the deterministic problem if the initial state is (x, z). It differs from the value function for the stochastic problem by a constant. Using constraints (4.10) and (4.11) to substitute for it and λt in the utility function, an indirect utility function U(x, z, d) is obtained. The value function, v(x, z), was computed by the method of successive approximations or value iteration. If vj(x, z) is the jth approximation, then subject to constraints (4.6)–(4.9). The initial approximation, v0(x, z), is that function which is identically zero. The function U is quadratic and the constraints are linear. Then, if vj is quadratic, υ j+1 must be quadratic. As υ 0 is trivially quadratic, all the υ j are quadratic and therefore easily computable. We found that the sequence of quadratic functions converged reasonably quickly.12 12 The limit of the sequence of value functions existed in every case and, as a function of z, was bounded from above, given x. This, along with the stability of the matrix A, is sufficient to ensure that this limit is the optimal value function and that the associated policy function is the optimal one (see [30]). © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors TIME TO BUILD AGGREGATE FLUCTUATIONS 71 1359 The next step is to determine the optimal inventory carry-over decision rule. It is the linear function yt+1=y(xt, zt, nt, sJt) which solves (4.12) subject to (4.6)–(4.9) and both nt and SJt given. Finally, the solution to the program where v2 is thevalue of maximization of (4.12), is determined. The linear functions SJt=s(xt, zt) and nt=n(xt, zt) which solve the above program are the optimal decision rules for new projects and labor supply. Because of the separation of estimation and control in our model, these decision rules can be used to determine the motion of the stochastic economy. In each period t, a conditional expectation, m0t, is formed on the basis of observations in previous periods. An indicator of the technology shock is observed, which is the sum of a permanent and a transitory component as well as an indicator shock. The conditional expectation, m1t, of the unobserved xt is computed according to equation (3.14), and SJt and nt are determined from (4.13) (4.14) where xt has been replaced by m 1t. Then the technology shock, t, is observed, which changes the conditional expectation of xt. From (3.16), this expectation is m2t, and the inventory carry-over is determined from (4.15) To summarize, the equilibrium process governing the evolution of our economy is given by (3.1)–(3.3), (3.6), (3.11)–(3.14), (3.16), (3.18), and (4.13)–(4.15). 5. TEST OF THE THEORY The test of the theory is whether there is a set of parameters for which the model’s co-movements for both the smoothed series and the deviations from the smoothed series are quantitatively consistent with the observed behavior of the corresponding series for the U.S. post-war economy. An added requirement is that the parameters selected not be inconsistent with relevant micro observations, including reported construction periods for new plants and cross-sectional observations on consumption and labor supply. The closeness of our specification of preferences and technology to those used in many applied studies facilitates such comparisons. The model has been rigged to yield the observations that smoothed output, investment, consumption, labor productivity, and capital stocks all vary roughly © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 72 FOUNDATIONS 1360 F.E.KYDLAND AND E.C.PRESCOTT proportionately while there is little change in employment (all variables are in perhousehold terms) when the technology parameter grows smoothly over time. These are just the steady state properties of the growth model with which we began. Quantitatively explaining the co-movements of the deviations is the test of the underlying theory. For want of better terminology, the deviations will be referred to as the cyclical components even though, with our integrated approach, there is no separation between factors determining a secular path and factors determining deviations from that path. The statistics to be explained are the covariations of the cyclical components. They are of interest because their behavior is stable and is so different from the corresponding covariations of the smoothed series. This is probably why many have sought separate explanations of the secular and cyclical movements. One cyclical observation is that, in percentage terms, investment varies three times as much as output does and consumption only half as much. In sharp contrast to the secular observations, variations in cyclical output are principally the result of variations in hours of employment per household and not in capital stocks or labor productivity. The latter observation is a difficult one to explain. Why does the consumption of market produced goods and the consumption of leisure move in opposite directions in the absence of any apparent large movement in the real wage over the so-called cycle? For our model, the real wage is proportional to labor’s productivity, so the crucial test is whether most of the variation in cyclical output arises from variations in employment rather than from variations in labor’s productivity. We chose not to test our model versus the less restrictive vector autoregressive model.13 This most likely would have resulted in the model being rejected, given the measurement problems and the abstract nature of the model. Our approach is to focus on certain statistics for which the noise introduced by approximations and measurement errors is likely to be small relative to the statistic. Failure of the theory to mimic the behavior of the post-war U.S. economy with respect to these stable statistics with high signal-noise ratios would be grounds for its rejection. Model Calibration There are two advantages of formulating the model as we did and then constructing an approximate model for which the equilibrium decision rules are linear. First, the specifications of preferences and technology are close to those used in many applied studies. This facilitates checks of reasonableness of many parameter values. Second, our approach facilitates the selection of parameter values for which the model steady-state values are near average values for the American economy during the period being explained. These two considerations 13 Sims [34] has estimated unrestricted aggregate vector autoregressive models. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors TIME TO BUILD AGGREGATE FLUCTUATIONS 73 1361 reduce dramatically the number of free parameters that will be varied when searching for a set that results in cyclical covariances near those observed. In explaining the covariances of the cyclical components, there are only seven free parameters, with the range of two of them being severely constrained a priori. Capital for our model reflects all tangible capital, including stocks of plant and equipment, consumer durables and housing. Consumption does not include the purchase of durables but does include the services from the stock of consumer durables. Different types of capital have different construction periods and patterns of resource requirements. The findings summarized in Section 2 suggest an average construction period of nearly two years for plants. Consumer durables, however, have much shorter average construction periods. Having but one type of capital, we assume, as a compromise, that four quarters are required, with one-fourth of the value put in place each quarter. Thus J=4 and 1=2 =3=4=0.25. Approximately ten per cent of national income account GNP is the capital consumption allowance and another ten per cent excise tax. To GNP should be added the depreciation of consumer durables which has the effect of increasing the share of output going to owners of capital. In 1976, compensation to employees plus proprietary income was approximately 64 per cent of GNP plus consumer durables depreciation less indirect business tax, while owners of capital received about 36 per cent. As labor share is , we set =0.64. Different types of capital depreciate more rapidly than others, with durables depreciating more rapidly than plant and housing, and land not depreciating at all. As a compromise, we set the depreciation rate equal to 10 per cent per year. We assume a subjective time discount rate of four per cent and abstract from growth. This implies a steady-state capital to annual output ratio of 2.4. Of total output 64 per cent is wages, 24 per cent depreciation, and 12 per cent return on capital which includes consumer durables. The remaining parameters of technology are average , which we normalize to one by measuring output in the appropriate units, and parameters and , which determine the shares of and substitution between inventories and capital. Inventories are about one-fourth of annual GNP so we require and to be such that k/y=10. A priori reasoning indicates the substitution opportunities between capital and inventory are small, suggesting that should be considerably larger than zero. We restricted it to be no less than two, but it is otherwise a free parameter in our search for a model to explain the cyclical covariances and autocovariances of aggregate variables. Given and the value of b1=y/k, is implied. From (4.1) it is For purposes of explaining the covariances of the percentage deviation from steady state values, is the only free parameter associated with technology. The steady state real interest rate r is related to the subjective time discount rate, and the risk aversion parameter, , by the equation r=+ (1-) (c/c), where c/c is the growth rate of per capita consumption. We have assumed is four per cent per year (one per cent per quarter). As the growth rate © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 74 FOUNDATIONS 1362 F.E.KYDLAND AND E.C.PRESCOTT of per capita consumption has been about two per cent and the real return on physical capital six to eight per cent, the risk aversion parameter, , is constrained to be between minus one and zero.14 The parameters 0 and which affect intertemporal substitutability of leisure will be treated as free parameters for we could find no estimate for them in the labor economics literature. As stated previously, the steady-state labor supply is independent of the productivity parameter . The remaining parameters are those specifying the process on t and the variance of the indicator. These three parameters are var(1), var(2), and var(3). Only two of these are free parameters, however. We restricted the sum of the three variances to be such that the estimate of the variance of cyclical output for the model equalled that of cyclical output for the U.S. economy during the sample period. In summary, the parameters that are estimated from the variance-covariance properties of the model are these variances plus the parameter determining substitutability of inventories and capital, the parameters 0 and determining intertemporal substitutability of leisure, and the risk aversion parameter . For each set of parameter values, means and standard deviations were computed for several statistics which summarize the serial correlation and covariance properties of the model. These numbers are compared with those of the actual U.S. data for the period 1950:1 to 1979:2 as reported in Hodrick and Prescott [18]. A set of parameter values is sought which fits the actual data well. Having only six degrees of freedom to explain the observed covariances imposes considerable discipline upon the analysis. The statistics reported in [18] are not the only way to quantitatively capture the co-movements of the deviations.15 This approach is simple, involves a minimum of judgment, and is robust to slowly changing demographic factors which affect growth, but are not the concern of this theory.16 In addition, these statistics are robust to most measurement errors, in contrast to, say, the correlations between the first differences of two series. It is important to compute the same statistics for the U.S. economy as for the model, that is, to use the same function of the data. This is what we do. A key part of our procedure is the computation of dynamic competitive equilibrium for each combination of parameter values. Because the conditional forecasting can be separated from control in this model, the dynamic equilibrium decision rules need only be computed for each new combination of the parameters 14 15 Estimates in [16] indicate is near zero. With the Hodrick-Prescott method, the smooth path {st} for each series {yt} minimized The deviations for series {yt} are {y t-s t}. The number of observations, T, was 118. The solution to the above program is a linear transformation of the data. Thus, the standard deviations and correlations reported are well-defined statistics. 16 See, for example, [11]. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors TIME TO BUILD AGGREGATE FLUCTUATIONS 75 1363 TABLE I a MODEL PARAMETERS a For parameters with a time dimension, the unit of time is a quarter of a year. , 0, , and . Similarly, the conditional expectations of the permanent and transitory shocks which enter the decision rules depend only on the variances of the three shocks and not upon the parameters of preferences and technology. For each set of parameter values the following statistics are computed: the autocorrelation of cyclical output for up to six periods, standard deviations of the cyclical variables of interest, and their correlations with cyclical output. In [18] the variables (except interest rates) are measured in logs while we use the levels rather than the logs. This is of consequence only in the measurement of amplitudes, so in order to make our results comparable to theirs, our standard deviations (except for interest rates) are divided by the steady states of the respective variables. One can then interpret the cyclical components essentially as percentage deviations as in [18]. The parameter values that yielded what we considered to be the best fit are reported in Table I. They were determined from a grid search over the free parameters. In the case of , we tried the values 2, 3, 4, and 5. The parameters 0 and were just constrained to be between zero and one. Only the values -1, -0.5, and -0.1 were considered for the risk aversion parameter . The last value is close to the limiting case of =0 which would correspond to the logarithmic utility function. Results All reported statistics refer to the cyclical components for both the model and the U.S. economy. Estimated autocorrelations of real output for our model along with sample values for the U.S. economy in the post-war period are reported in Table II. The fit is very good, particularly in light of the model’s simplicity. Table III contains means of standard deviations and correlations with output for the model’s variables. Table IV contains sample values of statistics for the post-war U.S. economy as reported in [18]. The variables in our model do not correspond perfectly to those available for the U.S. economy so care must be taken in making comparisons. A second problem is that there may be measurement errors that seriously bias the estimated correlations and standard deviations. A final problem is that the estimates for the U.S. economy are subject to sampling error. As a guide to the magnitude of this variability, © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 76 FOUNDATIONS 1364 F.E.KYDLAND AND E.C.PRESCOTT TABLE II a AUTOCORRELATIONS OF OUTPUT a The length of the sample period both for the model and for the U.S. economy is 118 quarters. we report the standard deviations of sample distributions for the model’s statistics which, like the estimates for the U.S. economy, use only 118 observations. These are the numbers in the parentheses in Tables II and III. The model is consistent with the large (percentage) variability in investment and low variability in consumption and their high correlations with real output. The model’s negative correlation between the capital stock and output is consistent with the data though its magnitude is somewhat smaller. Inventories for our model correspond to finished and nearly finished goods while the inventories in Table IV refer to goods in process as well. We added half TABLE III MODEL’S STANDARD DEVIATIONS AND CORRELATIONS WITH REAL OUTPUTa a b The length of the sample period both for the model and for the U.S. economy is 118 quarters. Measured in per cent. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors TIME TO BUILD AGGREGATE FLUCTUATIONS 77 1365 TABLE IV SAMPLE STANDARD DEVIATIONS AND CORRELATIONS WITH REAL OUTPUT U.S. ECONOMY 1950:1–1979:2 the value of uncompleted capital goods to the model’s inventory variable to obtain what we call inventories plus. This corresponds more closely to the U.S. inventory stock variable, with its standard deviation and correlation with real output being consistent with the U.S. data. In Table III we include results for the implicit real interest rate given by the expression rt=(u/ct)/( E(u/ct+l))-1. The expectation is conditional on the information known when the allocation between consumption and inventory carryover is made. The model displays more variability in hours than in productivity, but not by as much as the data show. In light of the difficulties in measuring output and, in particular, employment, we do not think this discrepancy is large. For example, all members of the household may not be equally productive, say due to differing stocks of human capital. If there is a greater representation in the work force of the less productive, for example less experienced youth, when output is high, hours would be overestimated. The effects of such errors would be to bias the variability of employment upwards. It also would bias the correlation between productivity and output downwards, which would result in the model being nearly consistent with the data. Measurement errors in employment that are independent of the cycle would have a similar effect on the correlation between output and productivity. Another possible explanation is the oversimplicity of the model. The shocks to technology, given our production function, are pure productivity shocks. Some shocks to technology alter the transformation between the consumption and investment goods. For example, investment tax credits, accelerated depreciation, and the like, have such effects, and so do some technological changes. Further, © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 78 FOUNDATIONS 1366 F.E.KYDLAND AND E.C.PRESCOTT some technological change may be embodied in new capital, and only after the capital becomes productive is there the increment to measured productivity. Such shocks induce variation in investment and employment without the variability in productivity. This is a question that warrants further research. We also examined lead and lag relationships and serial correlation properties of aggregate series other than output. We found that, both for the post-war U.S. economy and the model, consumption and non-inventory investment move contemporaneously with output and have serial correlation properties similar to output. Inventory and capital stocks for the model lag output, which also matches well with the data. Some of the inventory stock’s cross-serial correlations with output deviate significantly, however, from those for the U.S. economy. The one variable whose lead-lag relationship does not match with the data is productivity. For the U.S. economy it is a leading indicator, while there is no lead or lag in the model. This was not unexpected in view of our discussion above with regard to productivity. Thus, even though the overall fit of the model is very good, it is not surprising, given the level of abstraction, that there are elements of the fine structure of dynamics that it does not capture. The Smoothed Series The smoothed output series for the U.S. post-war data deviated significantly from the linear time trend. During the 118-quarter sample period this difference had two peaks and two troughs. The times between such local extremes were 30, 31, and 32 quarters, and the corresponding differences in values at adjacent extremes were 5.00, 7.25, and 5.90 per cent, respectively. These observations match well with the predictions of the model. The mean of the model’s sampling distribution for the number of peaks and troughs in a 118quarter period is 4.0—which is precisely the number observed. The mean of the number of quarters between extremes is 26.1 with standard deviation 9.7, and the mean of the vertical difference in the values at adjacent extremes is 5.0 with standard deviation 2.9. Thus, the smoothed output series for the U.S. economy is also consistent with the model. Sensitivity of Results to Parameter Selection With a couple of exceptions, the results were surprisingly insensitive to the values of the parameters. The fact that the covariations of the aggregate variables in the model are quite similar for broad ranges of many of the parameters suggests that, even though the parameters may differ across economies, the nature of business cycles can still be quite similar. We did find that most of the variation in technology had to come from its permanent component in order for the serial correlation properties of the model to be consistent with U.S. post-war data. We also found that the variance of the indicator shock could not be very large relative to the variance of the permanent © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors TIME TO BUILD AGGREGATE FLUCTUATIONS 79 1367 technology shock. This would have resulted in cyclical employment varying less than cyclical productivity which is inconsistent with the data. Of particular importance for the model is the dependence of current utility on past leisure choices which admits greater intertemporal substitution of leisure. The purpose of this specification is not to contribute to the persistence of output changes. If anything, it does just the opposite. This element of the model is crucial in making it consistent with the observation that cyclical employment fluctuates substantially more than productivity does. For the parameter values in Table I, the standard deviation of hours worked is 18 per cent greater than the deviation of productivity. The special case of 0=1 corresponds to a standard time-separable utility function. For this case, with the parameters otherwise the same as in Table I, the standard deviation of hours is 24 per cent less than the deviation of productivity. Importance of Time to Build Of particular interest is the sensitivity of our results to the specification of investment technology. The prominent alternative to our time-to-build technology is the adjustment-cost structure. If only one period is required for the construction of new productive capital, we can write the law of motion for the single capital good as kt+1=(1-)kt+st, where st is the amount of investment in productive capital in period t. We can then introduce cost of adjustment into the model by modifying the resource constraint (3.4) as follows: where the parameter is nonnegative. The model in Section 3 implied that the price of investment goods, it, relative to consumption goods, ct, must be one. This price will now of course generally not equal one, but our cost-of-adjustment formulation insures that it is one when net investment is zero. The magnitude of the adjustment cost can probably best be judged in terms of the effect it has on this relative price of investment goods which differs from one by the amount 2(st-kt). If, for example, the parameter is 0.5, and the economy is near its steady state, a one per cent increase in the relative price of the investment good would be associated with a four per cent increase in gross investment which is approximately one per cent of GNP. Even when the adjustment cost is of this small magnitude, the covariance properties of the model are grossly inconsistent with the U.S. data for the post-war period. In particular, most of the fluctuation of output in this model is caused by productivity changes rather than changes in work hours. The standard deviation of hours is 0.60, while the standard deviation of productivity is 1.29. This is just the opposite of what the U.S. data show. Further evidence of the failure of the cost-of-adjustment model is that, relative to the numbers reported in Table III for our model, the standard deviation is © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 80 FOUNDATIONS 1368 F.E.KYDLAND AND E.C.PRESCOTT nearly doubled for consumption and reduced by a factor of two for investment expenditures, making the amplitudes of these two output components much too close as compared with the data. In addition, the standard deviation of capital stock was reduced by more than one half. The results were even worse for larger values of . The extreme case of =0 corresponds to the special case of J=1 in our model. Thus, neither time to build nor cost of adjustment would be an element of the model. The biggest changes in the results for this version as compared with Table III are that the correlation between capital stock and output becomes positive and of sizable magnitude (0.43 if the parameters are otherwise the same as in Table I), and that the correlation between inventory stock and output becomes negative (-0.50 for our parameter values). Both of these correlations are inconsistent with the observations. Also, the persistence of movements in investment expenditures as measured by the autocorrelations was substantially reduced. For our model with multiple periods required to build new capital, the results are not overly sensitive to the number of periods assumed. With a three or five-quarter construction period instead of four, the fit is also good. 6. CONCLUDING COMMENTS A competitive equilibrium model was developed and used to explain the autocovariances of real output and the covariances of cyclical output with other aggregate economic time series for the post-war U.S. economy. The preferencetechnology environment used was the simplest one that explained the quantitative comovements and the serial correlation properties of output. These results indicate a surprisingly good fit in light of the model’s simplicity. A crucial element of the model that contributed to persistence of output movements was the time-to-build requirement.17 We experimented with adjustment costs, the standard method for introducing persistence (e.g., [4, 33]), and found that they were not a substitute for the time-to-build assumption in explaining the data.18 One problem was that, even with small adjustment costs, employment and investment fluctuations were too small and consumption fluctuations too large to match with the observations. There are several refinements which should improve the performance of the model. In particular, we conjecture that introducing as a decision variable the hours 17 Capital plays an important role in creating persistence in the analysis of Lucas [23] as well as in those of Blinder and Fischer [5] and Long and Plosser [22]. In [23] gradual diffusion of information also plays a crucial role. This is not the case in our model, however, as agents learn the value of the shock at the end of the period. Townsend [37] analyzes a model in which decision makers forecast the forecasts of others, which gives rise to confounding of laws of motion with forecasting problems, and results in persistence in capital stock and output movements. 18 An alternative way of obtaining persistence is the use of long-term staggered nominal wage contracts as in [35]. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors TIME TO BUILD AGGREGATE FLUCTUATIONS 81 1369 per week that productive capital is employed, with agents having preferences defined on hours worked per week, should help. Introducing more than a single type of productive capital, with different types requiring different periods for construction and having different patterns of resource requirement, is feasible. It would then be possible to distinguish between plant, equipment, housing, and consumer durables investments. This would also have the advantage of permitting the introduction of features of our tax system which affect transformation opportunities facing the economic agents (see, e.g., [14]). Another possible refinement is in the estimation procedure. But, in spite of the considerable advances recently made by Hansen and Sargent [15], further advances are needed before formal econometric methods can be fruitfully applied to testing this theory of aggregate fluctuations. Models such as the one considered in this paper could be used to predict the consequence of a particular policy rule upon the operating characteristics of the economy.19 As we estimate the preference-technology structure, our structural parameters will be invariant to the policy rule selected even though the behavioral equations are not. There are computational problems, however, associated with determining the equilibrium behavioral equations of the economy when feedback policy rules, that is, rules that depend on the aggregate state of the economy, are used. The competitive equilibrium, then, will not maximize the welfare of the standin consumer, so a particular maximization problem cannot be solved to find the equilibrium behavior of the economy. Instead, methods such as those developed in [20] to analyze policy rules in competitive environments will be needed. Carnegie-Mellon University and University of Minnesota Manuscript received January, 1981; revision received January, 1982. 19 Examples of such policy issues are described in [21]. See also Barro (e.g., [3]), who emphasizes the differences in effects of temporary and permanent changes in government expenditures. REFERENCES [1] ABEL, A.B.: “Empirical Investment Equations: An Integrated Framework,” in On The State of Macroeconomics, ed. by K.Brunner and A.H.Meltzer. Amsterdam: North-Holland, 1979. [2] ABOWD, J., AND O.ASHENFELTER: “Unemployment and Compensating Wage Differentials,” in Studies in Labor Markets, ed. by S.Rosen. Chicago: University of Chicago Press, 1981. [3] BARRO, R.J.: “Intertemporal Substitution and the Business Cycle,” in Supply Shocks, Incentives and National Wealth, ed. by K.Brunner and A.H.Meltzer. Amsterdan: North-Holland, 1981. [4] BLACK, F.: “General Equilibrium and Business Cycles,” Working Paper, Massachusetts Institute of Technology, revised November, 1979. [5] BLINDER, A.S., AND S.FISCHER: “Inventories, Rational Expectations, and the Business Cycle,” Journal of Monetary Economics, 8 (1981), 227–304. [6] BÖHM-BAWERK, E. VON: Positive Theory of Capital, trans. by W.Smart. London: Macmillan, 1891. [7] BRUNNER, K., AND A.H.MELTZER: “Money, Debt, and Economic Activity,” Journal of Political Economy, 80 (1972), 951–977. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 82 FOUNDATIONS 1370 F.E. KYDLAND AND E.C.PRESCOTT [8] BRUNNER, K., A. CUKIERMAN, AND A.H.MELTZER: “Stagflation, Persistent Unemployment and the Permanence of Economic Shocks,” Journal of Monetary Economics, 6(1980), 467–492. [9] CARLTON, D.W.: “Contracts, Price Rigidity, and Market Equilibrium,” Journal of Political Economy, 87(1979), 1034–1062. [10] CHRISTIANO, L.J.: “On the Accuracy of Linear Quadratic Approximations: An Example,” Working Paper, University of Chicago, September, 1981. [11] DENISON, E.F.: Accounting for United States Economic Growth 1929–1969. Washington, D.C.: Brookings Institution, 1974. [12] GROSSMAN, H.I.: “Aggregate Demand, Job Search, and Employment,” Journal of Political Economy, 81(1973), 1353– 1369. [13] HALL, R.E.: “Investment, Interest Rates, and the Effects of Stabilization Policies,” Brookings Papers on Economic Activity, 6(1977), 61–100. [14] HALL, R.E., AND D.W.JORGENSON: “Tax Policy and Investment Behavior,” American Economic Review, 57(1967), 391–414. [15] HANSEN, L.P., AND T.J.SARGENT: “Formulating and Estimating Dynamic Linear Rational Expectations Models,” Journal of Economic Dynamics and Control, 2(1980), 7–46. [16] HANSEN, L.P., AND K.J.SINGLETON: “Generalized Instrumental Variables Estimation of Nonlinear Rational Expectations Models,” Econometrica, 50(1982), 1269–1286. [17] HAYASHI, F.: “Tobin’s Marginal q and Average q: A Neoclassical Interpretation,” Econometrica, 50(1982), 213–224. [18] HODRICK, R.J., AND E.C.PRESCOTT: “Post-War U.S. Business Cycles: An Empirical Investigation,” Working Paper, Carnegie-Mellon University, revised November, 1980. [19] JORGENSON, D.W.: “Anticipations and Investment Behavior,” in The Brookings Quarterly Econometric Model of the United States, ed. by J.S.Duesenberry et al. Chicago: Rand McNally, 1965. [20] KYDLAND, F.E.: “Analysis and Policy in Competitive Models of Business Fluctuations,” Working Paper, CarnegieMellon University, revised April, 1981. [21] KYDLAND, F.E., AND E.C.PRESCOTT: “A Competitive Theory of Fluctuations and the Feasibility and Desirability of Stabilization Policy,” in Rational Expectations and Economic Policy, ed. by S.Fischer. Chicago: University of Chicago Press, 1980. [22] LONG, J.B., JR., AND C.I.PLOSSER: “Real Business Cycles,” Working Paper, University of Rochester, November, 1980. [23] LUCAS, R.E., JR.: “An Equilibrium Model of the Business Cycle,” Journal of Political Economy, 83(1975), 1113–1144. [24]——: “Understanding Business Cycles,” in Stabilization of the Domestic and International Economy, ed. by K.Brunner and A.H.Meltzer. New York: North-Holland, 1977. [25] LUCAS, R.E., JR, AND L.A.RAPPING: “Real Wages, Employment and Inflation,” Journal of Political Economy, 77(1969), 721–754. [26] MALKIEL, B.G., G.M.VON FURSTENBERG, AND H.S.WATSON: “Expectations, Tobin’s q, and Industry Investment,” Journal of Finance, 34(1979), 549–561. [27] MAYER, T.: “Plant and Equipment Lead Times,” Journal of Business, 33(1960), 127–132. [28] MOOD, A.M., AND F.A.GRAYBILL: Introduction to the Theory of Statistics, 2nd ed. New York: McGraw-Hill, 1963. [29] MUSSA, M.: “External and Internal Adjustment Costs and the Theory of Aggregate and Firm Investment,” Economica, 44(1977), 163–178. [30] PRESCOTT, E.C.: “A Note on Dynamic Programming with Unbounded Returns,” Working Paper, University of Minnesota, 1982. [31] PRESCOTT, E.C., AND R.MEHRA: “Recursive Competitive Equilibrium: The Case of Homogeneous Households,” Econometrica, 48(1980), 1365–1379. [32] SARGENT, T.J.: “Tobin’s q and the Rate of Investment in General Equilibrium,” in On the State of Macroeconomics, ed. by K.Brunner and A.H.Meltzer. Amsterdam: North-Holland, 1979. [33]——: Macroeconomic Theory. New York: Academic Press, 1979. [34] SIMS, C.A.: “Macroeconomics and Reality,” Econometrica, 48(1980), 1–48. [35] TAYLOR, J.B.: “Aggregate Dynamics and Staggered Contracts,” Journal of Political Economy, 88(1980), 1–23. [36] TOBIN, J.: “A General Equilibrium Approach to Monetary Theory,” Journal of Money, Credit, and Banking, 1(1969), 15–29. [37] TOWNSEND, R.M.: “Forecasting the Forecasts of Others,” Working Paper, Carnegie-Mellon University, revised August, 1981. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors CHAPTER 4 83 Edward C.Prescott Theory Ahead of Measurement Theory Ahead of Business Cycle Measurement* Edward C.Prescott Adviser Research Department Federal Reserve Bank of Minneapolis and Professor of Economics University of Minnesota Economists have long been puzzled by the observations that during peacetime industrial market economies display recurrent, large fluctuations in output and employment over relatively short time periods. Not uncommon are changes as large as 10 percent within only a couple of years. These observations are considered puzzling because the associated movements in labor’s marginal product are small. These observations should not be puzzling, for they are what standard economic theory predicts. For the United States, in fact, given people’s ability and willingness to intertemporally and intratemporally substitute consumption and leisure and given the nature of the changing production possibility set, it would be puzzling if the economy did not display these large fluctuations in output and employment with little associated fluctuations in the marginal product of labor. Moreover, standard theory also correctly predicts the amplitude of these fluctuations, their serial correlation properties, and the fact that the investment component of output is about six times as volatile as the consumption component. This perhaps surprising conclusion is the principal finding of a research program initiated by Kydland and me (1982) and extended by Kydland and me (1984), Hansen (1985a), and Bain (1985). We have computed the competitive equilibrium stochastic process for variants of the constant elasticity, stochastic growth model. The elasticities of substitution and the share parameters of the production and utility functions are restricted to those that generate the growth observations. The process governing the technology parameter is selected to be consistent with the measured technology changes for the American economy since the Korean War. We ask whether these artificial economies display fluctuations with statistical properties similar to those which the American economy has displayed in that period. They do.1 I view the growth model as a paradigm for macro analysis—analogous to the supply and demand construct of price theory. The elasticities of substitution and the share parameters of the growth model are analogous to the price and income elasticities of price theory. Whether or not this paradigm dominates, as I expect it will, is still an open question. But the early results indicate its power to organize our knowledge. The finding that when uncertainty in the rate of technological change is incorporated into the growth model it displays the business cycle *This paper was presented at a Carnegie-Rochester Conference on Public Policy and will appear in a volume of the conference proceedings. It appears here with the kind permission of Allan H.Meltzer, editor of that volume. The author thanks Finn E.Kydland for helpful discussions of the issues reviewed here, Gary D.Hansen for data series and some additional results for his growth economy, Lars G.M.Ljungqvist for expert research assistance, Bruce D.Smith and Allan H.Meltzer for comments on a preliminary draft, and the National Science Foundation and the Minneapolis Federal Reserve Bank for financial support. The views expressed herein are those of the author alone. 1 Others [Barro (1981) and Long and Plosser (1983), for example] have argued that these fluctuations are not inconsistent with competitive theory that abstracts from monetary factors. Our finding is much stronger: standard theory predicts that the economy will display the business cycle phenomena. 9 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 84 FOUNDATIONS phenomena was both dramatic and unanticipated. I was sure that the model could not do this without some features of the payment and credit technologies. The models constructed within this theoretical framework are necessarily highly abstract. Consequently, they are necessarily false, and statistical hypothesis testing will reject them. This does not imply, however, that nothing can be learned from such quantitative theoretical exercises. I think much has already been learned and confidently predict that much more will be learned as other features of the environment are introduced. Prime candidates for study are the effects of public finance elements, a foreign sector, and, of course, monetary factors. The research I review here is best viewed as a very promising beginning of a much larger research program. The Business Cycle Phenomena The use of the expression business cycle is unfortunate for two reasons. One is that it leads people to think in terms of a time series’ business cycle component which is to be explained independently of a growth component; our research has, instead, one unifying theory of both of these. The other reason I do not like to use the expression is that it is not accurate; some systems of low-order linear stochastic difference equations with a nonoscillatory deterministic part, and therefore no cycle, display key business cycle features. (See Slutzky 1927.) I thus do not refer to business cycles, but rather to business cycle phenomena, which are nothing more nor less than a certain set of statistical properties of a certain set of important aggregate time series. The question I and others have considered is, Do the stochastic difference equations that are the equilibrium laws of motion for the stochastic growth display the business cycle phenomena? More specifically, we follow Lucas (1977, p. 9) in defining the business cycle phenomena as the recurrent fluctuations of output about trend and the co-movements among other aggregate time series. Fluctuations are by definition deviations from some slowly varying path. Since this slowly varying path increases monotonically over time, we adopt the common practice of labeling it trend. This trend is neither a measure nor an estimate of the unconditional mean of some stochastic process. It is, rather, defined by the computational procedure used to fit the smooth curve through the data. If the business cycle facts were sensitive to the detrending procedure employed, there would be a problem. But the key facts are not sensitive to the procedure if the trend curve is smooth. Our curve-fitting method is to take the logarithms of variables and then select the trend path {τt} which minimizes the sum of the squared deviations from a given series {Yt} subject to the constraint that the sum of the squared second differences not be too large. This is The smaller is µ, the smoother is the trend path. If µ=0, the least squares linear time trend results. For all series, µ is picked so that the Lagrange multiplier of the constraint is 1600. This produces the right degree of smoothness in the fitted trend when the observation period is a quarter of a year. Thus, the sequence {τt} minimizes The first-order conditions of this minimization problem are linear in Yt and tt, so for every series, t=AY, where A is the same T×T matrix. The deviations from trend, also by definition, are Unless otherwise stated, these are the variables used in the computation of the statistics reported here for both the United States and the growth economies. An alternative interpretation of the procedure is that it is a high pass linear filter. The facts reported here are essentially the same if, rather than defining the deviations by Yd=(I-A}Y, we filtered the Y using a high pass band filter, eliminating all frequencies of 32 quarters or greater. An advantage of our procedure is that it deals better with the ends of the sample problem and does not require a stationary time series. To compare the behaviors of a stochastic growth economy and an actual economy, only identical statistics for the two economies are used. By definition, a statistic is a real valued function of the raw time series. Consequently, if a comparison is made, say, between the standard deviations of the deviations, the date t deviation for the growth economy must be the same function of the data generated by that model as the date t deviation for the American economy is of that economy’s data. Our definitions of the deviations satisfy this criterion. 10 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THEORY AHEAD OF MEASUREMENT 85 Edward C.Prescott Theory Ahead of Measurement Figure 1 plots the logs of actual and trend output for the U.S. economy during 1947–82, and Figure 2 the corresponding percentage deviations from trend of output and hours of market employment. Output and hours clearly move up and down together with nearly the same amplitudes. Table 1 contains the standard deviations and cross serial correlations of output and other aggregate time series for the American economy during 1954–82. Consumption appears less variable and investment more variable than output. Further, the average product of labor is procyclical but does not vary as much as output or hours. Figure 1 Actual and Trend Logs of U.S. Gross National Product Quarterly, 1947–82 The Growth Model This theory and its variants build on the neoclassical growth economy of Solow (1956) and Swan (1956). In the language of Lucas (1980, p. 696), the model is a “fully articulated, artificial economic system” that can be used to generate economic time series of a set of important econom c aggregates. The model assumes an aggregate production function with constant returns to scale, inputs labor n and capital k, and an output which can be allocated either to current consumption c or to investment x. If t denotes the date, f: R2→R the production function, and zt a technology parameter, then the production constraint is Source of basic data: Citicorp’s Citibase data bank Figure 2 Deviations From Trend of Gross National Product and Nonfarm Employee Hours in the United States Quarterly, 1947–82 xt+ctⱕztf(kt, nt) where xt, ct, kt, nt ⱖ0. The model further assumes that the services provided by a unit of capital decrease geometrically at a rate 0<δ<1: kt+1=(1-δ)kt+xt. Solow completes the specification of his economy by hypothesizing that some fraction 0<σ<1 of output is invested and the remaining fraction 1-σ consumed and that nt is a constant— say, n-—for all t. For this economy, the law of motion of capital condition on zt is Once the {zt} stochastic process is specified, the stochastic process governing capital and the other economic aggregates are determined and realizations of the stochastic process can be generated by a computer. This structure is far from adequate for the study of Source of basic data: Citicorp’s Citibase data bank the business cycle because in it neither employment nor the savings rate varies, when in fact they do. Being explicit about the economy, however, naturally leads to the question of what determines these variables, which are central to the cycle. That leads to the introduction of a stand-in household with some explicit preferences. If we 11 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 86 FOUNDATIONS Table 1 Cyclical Behavior of the U.S. Economy Deviations From Trend of Key Variables, 1954:1–1982:4 Source of basic data: Citicorp's Citibase data bank abstract from the labor supply decision and and ), the uncertainty (that is, standard form of the utility function is where β is the subjective time discount factor. The function u:R+→R is twice differentiable and concave. The commodity space for the deterministic version of this model is l∞, infinite sequences of uniformly bounded consumptions The theorems of Bewley (1972) could be applied to establish existence of a competitive equilibrium for this l ∞ commodity-space economy. That existence argument, however, does not provide an algorithm for computing the equilibria. An alternative approach is to use the competitive welfare theorems of Debreu (1954). Given local nonsaturation and no externalities, competitive equilibria are Pareto optima and, with some additional conditions that are satisfied for this economy, any Pareto optimum can be supported as a competitive equilibrium. Given a single agent and the convexity, there is a unique optimum and that optimum is the unique competitive equilibrium allocation. The advantage of this approach is that algorithms for computing solutions to concave programming problems can be used to find the competitive equilibrium allocation for this economy. Even with the savings decision endogenous, this economy has no fluctuations. As shown by Cass (1965) and Koopmans (1965), the competitive equilibrium path converges monotonically to a unique rest point or, if zt is growing exponentially, to a balanced growth path. There are multisector variants of this model 12 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THEORY AHEAD OF MEASUREMENT 87 Edward C.Prescott Theory Ahead of Measurement in which the equilibrium path oscillates. (See Benhabib and Nishimura 1985 and Marimon 1984.) But I know of no multisector model which has been restricted to match observed factor shares by sector, which has a value for β consistent with observed interest rates, and which displays oscillations. When uncertainty is introduced, the household’s objective is its expected discounted utility: The commodity vector is now indexed by the is history of shocks; that is, the commodity point. As Brock and Mirman (1972) show, if the {zt} are identically distributed random variables, an optimum to the social planner’s problem exists and the optimum is a stationary stochastic process with kt+1=g(kt, zt) and ct=c(kt, zt). As Lucas and Prescott (1971) show, for a class of economies that include this one, the social optimum is the unique competitive equilibrium allocation. They also show that for these homogeneous agent economies, the social optimum is also the unique sequence-of-markets equilibrium allocation. Consequently, there are equilibrium time-invariant functions for the wage wt=w(kt, zt) and the rental price of capital rt=r(kt, zt), where these prices are relative to the date t consumption good. Given these prices, the firm’s period t problem is subject to the output constraint ytⱕztf(kt, nt). The household’s problem is more complicated, for it must form expectations of future prices. If at is its capital stock, its problem is and given a0-k0. In forming expectations, a household knows the relation between the economy’s state (kt, zt) and prices, wt=w(kt, zt) and rt=r(kt, zt). Further, it knows the process governing the evolution of the per capita capital stock, a variable which, like prices, is taken as given. The elements needed to define a sequence-ofmarkets equilibrium are the firm’s policy functions y(kt, zt), n(kt, zt), and k(kt, zt); the household’s policy functions x(at, kt, zt) and c(at, kt, zt); a law of motion of per capita capital kt+1=g(kt, zt); and pricing functions w(kt, zt) and r(kt, zt). For equilibrium, then, • The firm’s policy functions must be optimal given the pricing functions. • The household’s policy functions must be optimal given the pricing functions and the law of motion of per capita capital. • Spot markets clear; that is, for all kt and zt (Note that the goods market must clear only when the representative household is truly representative, that is, when at=kt.) • Expectations are rational; that is, g(kt, zt)=(1-δ)kt+x(kt, kt, zt). This definition still holds if the household values productive time that is allocated to nonmarket activities. Such time will be called leisure and denoted lt. The productive time endowment is normalized to 1, and the household faces the constraints n t +l t ⱕ 1 for all t. In addition, leisure is introduced as an argument of the utility function, so the household’s objective becomes the maximization of Now leisure—and therefore employment—varies in equilibrium. The model needs one more modification: a relaxation of the assumption that the technology shocks zt are identically and independently distributed random variables. As will be documented, they are not so distributed. Rather, 13 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 88 FOUNDATIONS they display considerable serial correlation, with their first differences nearly serially uncorrelated. To introduce high persistence, we assume zt+1=zt+⑀t+1 where the {⑀t+1} are identically and independently distributed and is near 1. With this modification, the recursive sequence-of-markets equilibrium definition continues to apply. Using Data to Restrict the Growth Model Without additional restrictions on preferences and technology, a wide variety of equilibrium processes are consistent with the growth model. The beauty of this model is that both growth and micro observations can be used to determine its production and utility functions. When they are so used, there are not many free parameters that are specific to explaining the business cycle phenomena and that cannot be measured independently of those phenomena. The key parameters of the growth model are the intertemporal and intratemporal elasticities of substitution. As Lucas (1980, p. 712) emphasizes, “On these parameters, we have a wealth of inexpensively available data from census cohort information, from panel data describing the reactions of individual households to a variety of changing market conditions, and so forth.” To this list we add the secular growth observations which have the advantage of being experiments run by nature with large changes in relative prices and quantities and with idiosyncratic factors averaged out.2 A fundamental thesis of this line of inquiry is that the measures obtained from aggregate series and those from individual panel data must be consistent. After all, the former are just the aggregates of the latter. Secularly in the United States, capital and labor shares of output have been approximately constant, as has r, the rental price of capital. However, the nation’s real wage has increased greatly—more than 100 percent since the Korean War. For these results to hold, the model’s production function must be approximately Cobb-Douglas: The share parameter is equal to labor’s share, which has been about 64 percent in the postwar period, so =0.64. This number is smaller than that usually obtained because we include services of consumer durables as part of output. This alternative accounting both reduces labor’s share and makes it more nearly constant over the postwar period. The artificial economy has but one type of capital, and it depreciates at rate ␦. In fact, different types of capital depreciate at different rates, and the pattern of depreciation over the life of any physical asset is not constant. Kydland and I (1982, 1984) simply pick ␦=0.10. With this value and an annual real interest rate of 4 percent, the steadystate capital-annual output ratio is about 2.6. That matches the ratio for the U.S economy and also implies a steady-state investment share of output near the historically observed average Except for parameters determining the process on the technology shock, this completely specifies the technology of the simple growth model. A key growth observation which restricts the utility function is that leisure per capita lt has shown virtually no secular trend while, again, the real wage has increased steadily. This implies an elasticity of substitution between consumption ct and leisure lt near 1. Thus the utility function restricted to display both constant intertemporal and unit intratemporal elasticities of substitution is where 1/␥>0 is the elasticity of substituting between different date composite commodities This leaves ␥ and the subjective time discount factor β [or, equivalently, the subjective time discount rate (1/β)-1] to be determined. The steady-state interest rate is i=(1/β)-1+␥(c/c). As stated previously, the average annual real interest rate is about 4 percent, and the growth rate of per capita consumption c/c has averaged nearly 2 percent. The following studies help restrict γ. Tobin and Dolde (1971) find that a γ near 1.5 is needed to match the life cycle consumption patterns of individuals. Using individual portfolio observations, Friend and Blume (1975) estimate γ to be near 2. Using aggregate stock market and consumption data, Hansen and Singleton (1983) estimate γ to be near 1. Using international data, Kehoe (1984) also finds a modest curvature parameter ␥. All these observations make a strong case that ␥ is not too far from 1. Since the nature of fluctuations of the artificial economy is not very 2 See Solow 1970 for a nice summary of the growth observations. 14 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THEORY AHEAD OF MEASUREMENT 89 Edward C.Prescott Theory Ahead of Measurement sensitive to ␥, we simply set ␥ equal to 1. Taking the limit as ␥→1 yields u(ct, lt)=(1-) log ct+ log lt. This leaves β and still to be determined. Hansen (1985b) has found that growing economies—that is, those with z t having a multiplicative, geometrically growing factor (1+λ)t with λ>0—fluctuate in essentially the same way as economies for which λ=0. This justifies considering only the case λ=0. If λ=0, then the average interest rate approximately equals the subjective time discount rate.3 Therefore, we set β equal to 0.96 per year or 0.99 per quarter. The parameter is the leisure share parameter. Ghez and Becker (1975) find that the household allocates approximately one-third of its productive time to market activities and two-thirds to nonmarket activities. To be consistent with that, the model’s parameter must be near two-thirds. This is the value assumed in our business cycle studies. Eichenbaum, Hansen, and Singleton (1984) use aggregate data to estimate this share parameter , and they obtain a value near five-sixths. The difference between two-thirds and five-sixths is large in the business cycle context. With =2/3, the elasticity of labor supply with respect to a temporary change in the real wage is 2, while if =5/6, it is 5. This is because a 1 percent change in leisure implies a /(-1) percent change in hours of employment. We do not follow the Eichenbaum-HansenSingleton approach and treat as a free parameter because it would violate the principle that parameters cannot be specific to the phenomena being studied. What sort of science would economics be if micro studies used one share parameter and aggregate studies another? The Nature of the Technological Change One method of measuring technological change is to follow Solow (1957) and define it as the changes in output less the sum of the changes in labor’s input times labor share and the changes in capital’s input times capital share. Measuring variables in logs, this is the percentage change in the technology parameter of the Cobb-Douglas production function. For the U.S. economy between the third quarter of 1955 and the first quarter of 1984, the standard deviation of this change is 1.2 percent.4 The serial autocorrelations of these changes are 1=-0.21, 2=-0.06, 3=0.04, 4= 0.01, and 5=-0.05. To a first approximation, the process on the percentage change in the technology process is a random walk with drift plus some serially uncorrelated measurement error. This error produces the negative first-order serial correlation of the differences. Further evidence that the random walk model is not a bad approximation is based on yearly changes. For the quarterly random walk model, the standard deviation of this change is 6.63 times the standard deviation of the quarterly change. For the U.S. data, the annual change is only 5.64 times as large as the quarterly change. This, along with the negative first-order serial correlation, suggests that the standard deviation of the persistent part of the quarterly change is closer to 5.64/6.63= 0.85 than to 1.2 percent. Some further evidence is the change over four-quarter periods— that is, the change from a given quarter of one year to the same quarter of the next year. For the random walk model, the standard deviation of these changes is 2 times the standard deviation of the quarterly change. A reason that the standard deviation of change might be better measured this way is that the measurement noise introduced by seasonal factors is minimized. The estimate obtained in this way is 0.95 percent. To summarize, Solow growth accounting finds that the process on the technology parameter is highly persistent with the standard deviation of change being about 0.90.5 The Solow estimate of the standard deviation of technological change is surely an overstatement of the variability of that parameter. There undoubtedly are nonnegligible errors in measuring the inputs. Since the capital input varies slowly and its share is small, the most serious measurement problem is with the labor input. Fortunately there are two independent measures of the aggregate labor input, one constructed from a survey of employers and the other from a survey of households. Under the assumption of orthogonality of their measurement errors, a reasonable estimate of the variance of the change in hours is the covariance between the changes in the two series. Since the household survey is not used to estimate aggregate output, I use the covariance between the changes in household hours and output as an estimate of the covariance 3 Actually, the average interest rate is slightly lower because of risk premia. Given the value of ␥ and the amount of uncertainty, the average premium is only a fraction of a percent. See Mehra and Prescott 1985 for further details. 4 I use Hansen’s (1984) human capital-weighted, household hour series. The capital stock and GNP series are from Citicorp’s Citibase data bank. 5 The process zt+1=.9zt+et+1 is, like the random walk process, highly persistent. Kydland and I find that it and the random walk result in essentially the same fluctuations. 15 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 90 FOUNDATIONS between aggregate hours and output. Still using a share parameter of =0.75, my estimate of the standard deviation of the percentage change in zt is the square root of ˆ cov(Δhˆ 1 , Δŷ)+ 2 cov(Δhˆ 1, Δhˆ 2), var(Δy)-2 where the caret ( ˆ ) denotes a measured value. For the sample period my estimate is 0.763 percent. This is probably a better estimate than the one which ignores measurement error. Still, my estimate might under- or overstate the variance of technological change. For example, the measurement of output might include significant errors. Perhaps measurement procedures result in some smoothing of the series. This would reduce the variability of the change in output and might reduce the covariance between measured hours and output. Another possibility is that changes in hours are associated with corresponding changes in capital’s utilization rate. If so, the Solow approach is inappropriate for measuring the technology shocks. To check whether this is a problem, I varied and found that = 0.85 yields the smallest estimate, 0.759, as opposed to 0.763 for =0.75. This suggests that my estimate is not at all sensitive to variations in capital utilization rates. To summarize, there is overwhelming evidence that technological shocks are highly persistent. But tying down the standard deviation of the technology change shocks is difficult. I estimate it as 0.763. It could very well be larger or smaller, though, given the accuracy of the measurements. The Statistical Behavior of the Growth Models Theory provides an equilibrium stochastic process for the growth economy studied. Our approach has been to document the similarities and differences between the statistical properties of data generated by this stochastic process and the statistical properties of American time series data. An alternative approach is to compare the paths of the growth model if the technological parameters {zt} were those experienced by the U.S. economy. We did not attempt this because theory’s predictions of paths, unlike its predictions of the statistical properties, are sensitive to what Learner (1983, p. 43) calls “whimsical” modeling assumptions. Another nontrivial problem is that the errors in measuring the innovations in the zt process are as large as the innovations themselves. The Basic Growth Model With the standard deviation of the technology shock equal to 0.763, theory implies that the standard deviation of output will be 1.48 percent. In fact, it is 1.76 percent for the postKorean War American economy. For the output of the artificial economy to be as variable as that, the variance of the shock must be 1.0, significantly larger than the estimate. The most important deviation from theory is the relative volatility of hours and output. Figure 3 plots a realization of the output and employment deviations from trend for the basic growth economy. A comparison of Figures 2 and 3 demonstrates clearly that, for the American economy, hours in fact vary much more than the basic growth model predicts. For the artificial economy, hours fluctuate 52 percent as much as output, whereas for the American economy, the ratio is 0.95. This difference appears too large to be a result of errors in measuring aggregate hours and output. The Kydland-Prescott Economy Kydland and I (1982, 1984) have modified the growth model in two important respects. First, we assume that a distributed lag of leisure and the market-produced good combine to produce the composite commodity good valued by the household. In particular, where ␣i+1/␣i=1- for i=1, 2, …and Figure 3 Deviations From Trend of GNP and Hours Worked in the Basic Growth Economy 16 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THEORY AHEAD OF MEASUREMENT 91 Edward C.Prescott Theory Ahead of Measurement Table 2 Cyclical Behavior of the Kydland-Prescott Economy* * These are the means of 20 simulations, each of which was 116 periods long. The numbers in parentheses are standard errors. Source: Kydland and Prescott 1984 Kydland (1983) provides justification for this preference ordering based on an unmeasured, household-specific capital stock that, like ct and lt, is an input in the production of the composite commodity. The economy studied has ␣0=0.5 and =0.1. This increases the variability of hours. The second modification is to permit the workweek of capital to vary proportionally to the workweek of the household. For this economy, increases in hours do not reduce the marginal product of labor as much, so hours fluctuate more in response to technology shocks of a given size. The statistical properties of the fluctuations for this economy are reported in Table 2. As is clear there, hours are now about 70 percent as variable as output. This eliminates much of the discrepancy between theory and measurement. If the standard deviation of the technology shock is 0.72 percent, then fluctuations in the output of this artificial economy are as large as those experienced in the U.S. economy. A comparison of Tables 1 and 2 shows that the Kydland-Prescott economy displays the business cycle phenomena. It does not quite demonstrate, however, that there would be a puzzle if the economy did not display the business cycle phenomena. That is because the parameters ␣0 and have not been well tied down by micro observations.6 Better measures of these parameters could either increase or decrease significantly the amount of the fluctuations accounted for by the uncertainty in the technological change. 6 Hotz, Kydland, and Sedlacek (1985) use annual panel data to estimate ␣0 and and obtain estimates near the Kydland-Prescott assumed values. 17 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 92 FOUNDATIONS The Hansen Indivisible Labor Economy Labor economists have estimated labor supply elasticities and found them to be small for fulltime prime-age males. (See, for example, Ashenfelter 1984.) Heckman (1984), however, finds that when movements between employment and nonemployment are considered and secondary workers are included, elasticities of labor supply are much larger. He also finds that most of the variation in aggregate hours arises from variation in the number employed rather than in the hours worked per employed person. These are the observations that led Hansen (1985a) to explore the implication of introducing labor indivisibilities into the growth model. As shown by Rogerson (1984), if the household’s consumption possibility set has nonconvexities associated with the mapping from hours of market production activities to units of labor services, there will be variations in the number employed rather than in the hours of work per employed person. In addition, the aggregate elasticity of labor supply will be much larger than the elasticity of those whose behavior is being aggregated. In this case aggregation matters, and matters greatly. There certainly are important nonconvexities in the mapping from hours of market activities to units of labor services provided. Probably the most important nonconvexity arises from the considerable amount of time required for commuting. Other features of the environment that would make full-time workers more than twice as productive as otherwise similar half-time workers are not hard to imagine. The fact that part-time workers typically are paid less per hour than full-time workers with similar human capital endowments is consistent with the existence of important nonconvexities. Hansen (1985a) restricts -each identical household to either work h hours or be unemployed. His relation is as depicted by the horizontal lines in Figure 4. This assumption is not as extreme as it appears. If the relation were as depicted by the curved line, the behavior of the economy would be the same. The key property is an initial convex region followed by a concave region in the mapping from hours of market activity to units of labor service. Figure 4 Relation Between Time Allocated to Market Activity and Labor Service With this modification, lotteries that specify the probability of employment are traded along with market-produced goods and capital services. As before, the utility function of each individual is u(c, l)=(1/3) log c+(2/3) log l. otherwise, If an individual works, l=1. Consequently, if is the probability of employment, an individual’s expected utility is Given that per capita consumption is c- and per capita hours of employment n-, average utility over the population is maximized by for all individuals. If l , which equals setting 1–h , denotes per capita leisure, then maximum per capita utility is This is the utility function which rationalizes the per capita consumption and leisure choices if each person’s leisure is constrained to be either or 1. The aggregate intertemporal elasticity of substitution between different date leisures is infinity independent of the value of the elasticity for the individual (in the range where not all are employed). Hansen (1985a) finds that if the technology shock standard deviation is 0.71, then fluctuations in output for his economy are as large as those for the American economy. Further, variability 18 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THEORY AHEAD OF MEASUREMENT 93 Edward C.Prescott Theory Ahead of Measurement in hours is 77 percent as large as variability in output. Figure 5 shows that aggregate hours and output for his economy fluctuate together with nearly the same amplitude. These theoretical findings are the basis for my statement in the introduction that there would be a puzzle if the economy did not display the business cycle phenomena. Empirical Labor Elasticity One important empirical implication of a shockto-technology theory of fluctuations is that the empirical labor elasticity of output is significantly larger than the true elasticity, which for the CobbDouglas production function is the labor share parameter. To see why, note that the capital stock varies little cyclically and is nearly uncorrelated with output. Consequently, the deviations almost satisfy yt=ht+zt where yt is output, ht hours, and zt the technology shock. The empirical elasticity is =cov(ht, yt)/var(ht) which, because of the positive correlation between ht and zt, is considerably larger than the model’s , which is 0.64. For the basic, Kydland-Prescott, and Hansen growth economies, the values of are 1.9, 1.4, and 1.3, respectively. Because of measurement errors, the empirical elasticity for the American economy is not wellestimated by simply computing the ratio of the covariance between hours and output and dividing by the variance of hours. The procedure I use is based on the following probability model: where the caret ( ˆ ) denotes a measured value. The ⑀it are measurement errors. Here, the hˆ1t measure of hours uses the employer survey data while the hˆ2t measure uses the household survey data. Since these are independent measures, a maintained hypothesis is that ⑀2t and ⑀3t are orthogonal. With this assumption, a reasonable Figure 5 Deviations From Trend of GNP and Hours Worked in Hansen’s Indivisible Labor Economy Source Gary D Hansen. Department of Economics. University of California Santa Barbara estimate of var(ht) is the sample covariance between hˆ1t and hˆ2t. Insofar as the measurement of output has small variance or ⑀1t is uncorrelated with the hours measurement errors or both, the covariance between measured output and either measured hours series is a reasonable estimate of the covariance between output and hours. These two covariances are 2.231×10-4 and 2.244×10-4 for the sample period, and I take the average as my estimate of cov(ht, yt) for the American economy. My estimate of the empirical labor elasticity of output is This number is considerably greater than labor’s share, which is about 0.70 when services of consumer durables are not included as part of output. This number strongly supports the importance of technological shocks in accounting for business cycle fluctuations. Nevertheless, the number is smaller than those for the KydlandPrescott and Hansen growth economies. One possible reason for the difference between the U.S. economy and the growth model empirical labor elasticities of output is cyclical measurement errors in output. A sizable part of the investment component of output is hard to measure and therefore not included in 19 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 94 FOUNDATIONS the U.S. National Product Accounts measure of output, the gross national product (GNP). In particular, a firm’s major maintenance expenditures, research and development expenditures, and investments in human capital are not included in GNP. In good times—namely, when output is above trend—firms may be more likely to undertake major repairs of a not fully depreciated asset, such as replacing the roof of a 30-year-old building which has a tax life of 35 years. Such an expenditure is counted as maintenance and therefore not included in GNP even though the new roof will provide productive services for many years. The incentive for firms to do this is tax savings: by expensing an investment rather than capitalizing it, current tax liabilities are reduced. Before 1984, when a railroad replaced its 90-pound rails, the expenditure was treated as a maintenance expense rather than an investment expenditure. If these and other types of unmeasured investment fluctuate in percentage terms more than output, as do all the measured investment components, the volatility of GNP is larger than measured. We do know that investment in rails was highly procyclical and volatile in the postwar period. A careful study is needed to determine whether the correction for currently unmeasured investment is small or large. Another reason to expect the American economy’s labor elasticity to be less than the model’s is that the model shocks are perfectly neutral with respect to the consumption and investment good transformation. Persistent shocks which alter the product transformation frontier between these goods would cause variation in output and employment but not in the productivity parameters. For fluctuations so induced, the empirical labor elasticity of output would be the true elasticity. Similarly, relatively permanent changes in the taxing of capital—such as altering depreciation rates, the corporate income tax rate, or the investment tax credit rate—would all result in fluctuations in output and employment but not in the productivity parameters. A final reason for actual labor elasticity to be less than the model’s is the way imports are measured. An increase in the price of imported oil, that is, an increase in the quantity of output that must be sacrificed for a given unit of that input, has no effect on measured productivity. From the point of view of the growth model, however, an oil price increase is a negative technology shock because it results in less output, net of the exports used to pay for the imported oil, available for domestic consumption and investment. Theory predicts that such shocks will induce variations in employment and output, even though they have no effect on the aggregate production function. Therefore, insofar as they are important, they reduce the empirical labor elasticity of output. Extensions The growth model has been extended to provide a better representation of the technology. Kydland and I (1982) have introduced a technology with more than one construction period for new production capacity.7 We have also introduced inventory as a factor of production. This improves the match between the model’s serial correlation properties and the U.S. postwar data, but has little effect on the other statistics. Kydland (1984) has introduced heterogeneity of labor and found that if there are transfers from high human capital people to low human capital people, theory implies that hours of the low fluctuate more than hours of the high. It also implies a lower empirical labor elasticity of output than the homogeneous household model. Bain (1985) has studied an economy that is richer in sectoral detail. His model has manufacturing, retailing, and service-producing sectors. A key feature of the technology is that production and distribution occur sequentially. Thus there are two types of inventories—those of manufacturers’ finished goods and those of final goods available for sale. With this richer detail, theory implies that different components of aggregate inventories behave in different ways, as seen in the data. It also implies that production is more volatile than final sales, an observation considered anomalous since inventories can be used to smooth production. (See, for example, Blinder 1984.) Much has been done. But much more remains to be explored. For example, public finance considerations could be introduced and theory used to predict their implications. As mentioned above, factors which affect the rental price of capital affect employment and output, and the nature of the tax system affects the rental price of capital. Theory could be used to predict the effect of temporary increases in government expenditures such as those in the early 1950s when defense expenditures increased from less than 5 to more than 13 percent of GNP. Theory of this type could also be used to predict the effect of termsof-trade shocks. An implication of such an exercise most likely will be that economies with persistent terms-of-trade shocks fluctuate 7 Altug (1983) has introduced two types of capital with different gestation periods. Using formal econometric methods, she finds evidence that the model’s fit is improved if plant and equipment investment are not aggregated. 20 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THEORY AHEAD OF MEASUREMENT 95 Edward C.Prescott Theory Ahead of Measurement differently than economies with transitory shocks. If so, this prediction can be tested against the observations. Another interesting extension would be to explicitly model household production. This production often involves two people, with one specializing in market production and the other specializing in household production while having intermittent or part-time market employment. The fact that, cyclically, the employment of secondary wage earners is much more volatile than that of primary wage earners might be explained. A final example of an interesting and not yet answered question is, How would the behavior of the Hansen indivisible labor economy change if agents did not have access to a technology to insure against random unemployment and instead had to self-insure against unemployment by holding liquid assets? In such an economy, unlike Hansen’s, people would not be happy when unemployed. Their gain of more leisure would be more than offset by their loss as an insurer. Answering this question is not straightforward, because new tools for computing equilibria are needed. Summary and Policy Implications Economic theory implies that, given the nature of the shocks to technology and people’s willingness and ability to intertemporally and intratemporally substitute, the economy will display fluctuations like those the U.S. economy displays. Theory predicts fluctuations in output of 5 percent and more from trend, with most of the fluctuation accounted for by variations in employment and virtually all the rest by the stochastic technology parameter. Theory predicts investment will be three or more times as volatile as output and consumption half as volatile. Theory predicts that deviations will display high serial correlation. In other words, theory predicts what is observed. Indeed, if the economy did not display the business cycle phenomena, there would be a puzzle. The match between theory and observation is excellent, but far from perfect. The key deviation is that the empirical labor elasticity of output is less than predicted by theory. An important part of this deviation could very well disappear if the economic variables were measured more in conformity with theory. That is why I argue that theory is now ahead of business cycle measurement and theory should be used to obtain better measures of the key economic time series. Even with better measurement, there will likely be significant deviations from theory which can direct subsequent theoretical research. This feedback between theory and measurement is the way mature, quantitative sciences advance. The policy implication of this research is that costly efforts at stabilization are likely to be counterproductive. Economic fluctuations are optimal responses to uncertainty in the rate of technological change. However, this does not imply that the amount of technological change is optimal or invariant to policy. The average rate of technological change varies much both over time within a country and across national economies. What is needed is an understanding of the factors that determine the average rate at which technology advances. Such a theory surely will depend on the institutional arrangements societies adopt. If policies adopted to stabilize the economy reduce the average rate of technological change, then stabilization policy is costly. To summarize, attention should be focused not on fluctuations in output but rather on determinants of the average rate of technological advance. 21 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 96 FOUNDATIONS References Altug, S. 1983. Gestation lags and the business cycle. Working paper, Carnegie-Mellon University. Ashenfelter, O. 1984. Macroeconomic analyses and microeconomic analyses of labor supply. In Essays on macroeconomic implications of financial and labor markets and political processes, ed. K.Brunner and A.H.Meltzer. Carnegie-Rochester Conference Series on Public Policy 21:117–55. Amsterdam: North-Holland. Bain, I.R.M. 1985. A theory of the cyclical movements of inventory stocks. Ph.D. dissertation, University of Minnesota. Barro, R.J. 1981. Intertemporal substitution and the business cycle. In Supply shocks, incentives and national wealth, ed. K.Brunner and A.H.Meltzer. Carnegie-Rochester Conference Series on Public Policy 14:237–68. Amsterdam: North-Holland. Benhabib, J., and Nishimura, K. 1985. Competitive equilibrium cycles. Journal of Economic Theory 35:284– 306. Bewley, T.F. 1972. Existence of equilibria in economies with infinitely many commodities. Journal of Economic Theory 4:514–40. Blinder, A.S. 1984. Can the production smoothing model of inventory behavior be saved? Working paper, Princeton University. Brock, W.A., and Mirman, L.J. 1972. Optimal economic growth and uncertainty: The discounted case. Journal of Economic Theory 4: 479–513. Cass, D. 1965. Optimum growth in an aggregative model of capital accumulation. Review of Economic Studies 32:233–40. Debreu, G. 1954. Valuation equilibrium and Pareto optimum. Proceedings of the National Academy of Science 70:558–92. Eichenbaum, M.S.; Hansen, L.P.; and Singleton, K.S. 1984. A time series analysis of representative agent models of consumption and leisure choice under uncertainty. Working paper, Carnegie-Mellon University. Friend, I., and Blume, M.E. 1975. The demand for risky assets. American Economic Review 65:900–22. Ghez, G.R., and Becker, G.S. 1975. The allocation of time and goods over the life cycle. New York: National Bureau of Economic Research. Hansen, G.D. 1984. Fluctuations in total hours worked: A study using efficiency units. Working paper, University of Minnesota. ——. 1985a. Indivisible labor and the business cycle. Journal of Monetary Economics 16:309–27. ——. 1985b. Growth and fluctuations. Working paper, University of California, Santa Barbara. Hansen, L.P., and Singleton, K.J. 1983. Stochastic consumption, risk aversion, and the temporal behavior of asset returns. Journal of Political Economy 91:249–65. Heckman, J. 1984. Comments on the Ashenfelter and Kydland papers. In Essays on macroeconomic implications of financial and labor markets and political processes, ed. K.Brunner and A.H.Meltzer. Carnegie-Rochester Conference Series on Public Policy 21:209–24. Amsterdam: North-Holland. Hotz, V.S.; Kydland, F.E.; and Sedlacek, G.L. 1985. Intertemporal preferences and labor supply. Working paper, Carnegie-Mellon University. Kehoe, P.J. 1984. Dynamics of the current account: Theoretical and empirical analysis. Working paper, Harvard University. Koopmans, T.C. 1965. On the concept of optimal economic growth. In The econometric approach to development planning. Chicago: Rand-McNally. Kydland, F.E. 1983. Nonseparable utility and labor supply. Working paper, Hoover Institution. ——. 1984. Labor-force heterogeneity and the business cycle. In Essays on macroeconomic implications of financial and labor markets and political processes, ed. K.Brunner and A.H.Meltzer. Carnegie-Rochester Conference Series on Public Policy 21:173–208. Amsterdam: NorthHolland. Kydland, F.E., and Prescott, E.C. 1982. Time to build and aggregate fluctuations. Econometrica 50–70. ——. 1984. The workweek of capital and labor. Research Department Working Paper 267, Federal Reserve Bank of Minneapolis. Leamer, E.E. 1983. Let’s take the con out of econometrics. American Economic Review 73:31–43. Long, J.B., and Plosser, C.I. 1983. Real business cycles. Journal of Political Economy 91:39–69. Lucas, R.E., Jr. 1977. Understanding business cycles. In Stabilization of the domestic and international economy, ed. K.Brunner and A.H.Meltzer. Carnegie-Rochester Conference Series on Public Policy 5:7–29. Amsterdam: North-Holland. ——. 1980. Methods and problems in business cycle theory. Journal of Money, Credit and Banking 12 : 696–715. Reprinted in Studies in business-cycle theory, pp. 271–96. Cambridge, Mass.: MIT Press, 1981. Lucas, R.E., Jr., and Prescott, E.C. 1971. Investment under uncertainty. Econometrica 39:659–81. Marimon, R. 1984. General equilibrium and growth under uncertainty: The turnpike property. Discussion Paper 624. Northwestern University, Center for Mathematical Studies in Economics and Management Science. Mehra, R. and Prescott, E.C. 1985. The equity premium: A puzzle. Journal of Monetary Economics 15:145–61. Rogerson, R.D. 1984. Indivisible labor, lotteries and equilibrium. In Topics in the theory of labor markets, chap. 1. Ph.D. dissertation, University of Minnesota. Slutzky, E. 1927. The summation of random causes as the source of cyclic processes. In Problems of Economic Conditions, ed. Conjuncture Institute, Moskva (Moskow), vol. 3, no. 1. Revised English version, 1937, in Econometrica 5:105–46. Solow, R.M. 1956. A contribution to the theory of economic growth. Quarterly Journal of Economics 70:65–94. ——. 1970. Growth theory. Oxford: Oxford University Press. Swan, T.W. 1956. Economic growth and capital accumulation. Economic Record 32:334–61. Tobin, J., and Dolde, W. 1971. Wealth, liquidity and consumption. In Consumer spending and monetary policy: The linkages. Monetary Conference Series 5:99–146. Boston: Federal Reserve Bank of Boston. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors CHAPTER 5 97 Federal Reserve Bank of Minneapolis Quarterly Review Fall 1986 Some Skeptical Observations on Real Business Cycle Theory* Lawrence H.Summers Professor of Economics Harvard University and Research Associate National Bureau of Economic Research The increasing ascendancy of real business cycle theories of various stripes, with their common view that the economy is best modeled as a floating Walrasian equilibrium, buffeted by productivity shocks, is indicative of the depths of the divisions separating academic macroeconomists. These theories deny propositions thought self-evident by many academic macroeconomists and all of those involved in forecasting and controlling the economy on a day-to-day basis. They assert that monetary policies have no effect on real activity, that fiscal policies influence the economy only through their incentive effects, and that economic fluctuations are caused entirely by supply rather than demand shocks. If these theories are correct, they imply that the macroeconomics developed in the wake of the Keynesian Revolution is well confined to the ashbin of history. And they suggest that most of the work of contemporary macroeconomists is worth little more than that of those pursuing astrological science. According to the views espoused by enthusiastic proponents of real business cycle theories, astrology and Keynesian economics are in many ways similar: both lack scientific support, both are premised on the relevance of variables that are in fact irrelevant, both are built on a superstructure of nonoperational and ill-defined concepts, and both are harmless only when they are ineffectual. The appearance of Ed Prescott’s stimulating paper, “Theory Ahead of Business Cycle Measurement,” affords an opportunity to assess the current state of real business cycle theory and to consider its prospects as a foundation for macroeconomic analysis. Prescott’s paper is brilliant in highlighting the appeal of real business cycle theories and making clear the assumptions they require. But he does not make much effort at caution in judging the potential of the real business cycle paradigm. He writes that “if the economy did not display the business cycle phenomena, there would be a puzzle,” characterizes without qualification economic fluctuations as “optimal responses to uncertainty in the rate of technological change,” and offers the policy advice that “costly efforts at stabilization are likely to be counter-productive.” Prescott’s interpretation of his title is revealing of his commitment to his theory. He does not interpret the phrase theory ahead of measurement to mean that we lack the data or measurements necessary to test his theory. Rather, he means that measurement techniques have not yet progressed to the point where they fully corroborate his theory. Thus, Prescott speaks of the key deviation of observation from theory as follows: “An important part of this deviation could very well disappear if the economic variables were measured more in conformity with theory. That is why I argue that theory is now ahead of business cycle measurement….” The claims of real business cycle theorists deserve serious assessment, especially given their *An earlier version of these remarks was presented at the July 25, 1986, meeting of the National Bureau of Economic Research Economic Fluctuations Group. 23 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 98 FOUNDATIONS source and their increasing influence within the economics profession. Let me follow Prescott in being blunt. My view is that real business cycle models of the type urged on us by Prescott have nothing to do with the business cycle phenomena observed in the United States or other capitalist economies. Nothing in Prescott’s papers or those he references is convincing evidence to the contrary. Before turning to the argument Prescott presents, let me offer one lesson from the history of science. Extremely bad theories can predict remarkably well. Ptolemaic astronomy guided ships and scheduled harvests for two centuries. It provided extremely accurate predictions regarding a host of celestial phenomena. And to those who developed it, the idea that the earth was at the center seemed an absolutely natural starting place for a theory. So, too, Lamarckian biology, with its emphasis on the inheritance of acquired characteristics, successfully predicted much of what was observed in studies of animals and plants. Many theories can approximately mimic any given set of facts; that one theory can does not mean that it is even close to right. Prescott’s argument takes the form of the construction of an artificial economy which mimics many of the properties of actual economies. The close coincidence of his model economy and the actual economy leads him to conclude that the model economy is a reasonable if abstract representation of the actual economy. This claim is bolstered by the argument that the model economy is not constructed to fit cyclical facts but is parameterized on the basis of microeconomic information and the economy’s long-run properties. Prescott’s argument is unpersuasive at four levels. Are the Parameters Right? First, Prescott’s claim to have parameterized the model on the basis of well-established microeconomic and long-run information is not sustainable. As one example, consider a parameter which Prescott identifies as being important in determining the properties of the model, the share of household time devoted to market activities. He claims that is one-third. Data on its average value over the last century indicate, as Martin Eichenbaum, Lars Hansen, and Kenneth Singleton (1986) have noted, an average value of one-sixth over the past 30 years. This seems right— a little more than half the adult population works, and those who work work about a quarter of the time. I am unable to find evidence supporting Prescott’s one-third figure in the cited book by Gilbert Ghez and Gary Becker (1975). To take another example, Prescott takes the average real interest rate to be 4 percent. Over the 30-year period he studies, it in fact averaged only about 1 percent. This list of model parameters chosen somewhat arbitrarily could be easily extended. A more fundamental problem lies in Prescott’s assumption about the intertemporal elasticity of substitution in labor supply. He cites no direct microeconomic evidence on this parameter, which is central to his model of cyclical fluctuations. Nor does he refer to any aggregate evidence on it. Rather, he relies on a rather selective reading of the evidence on the intertemporal elasticity of substitution in consumption in evaluating the labor supply elasticity. My own reading is that essentially all the available evidence suggests only a minimal response of labor to transitory wage changes. Many studies (including Altonji 1982; Mankiw, Rotemberg, and Summers 1985; and Eichenbaum, Hansen, and Singleton 1986) suggest that the intertemporal substitution model cannot account at either the micro or the macro level for fluctuations in labor supply. Prescott is fond of parameterizing models based on long-run information. Japan has for 30 years enjoyed real wage growth at a rate four times the U.S. rate, close to 8 percent. His utility function would predict that such rapid real wage growth would lead to a much lower level of labor supply by the representative consumer. I am not aware that this pattern is observed in the data. Nor am I aware of data suggesting that age/hours profiles are steeper in professions like medicine or law, where salaries rise rapidly with age. Prescott’s growth model is not an inconceivable representation of reality. But to claim that its parameters are securely tied down by growth and micro observations seems to me a gross overstatement. The image of a big loose tent flapping in the wind comes to mind. Where Are the Shocks? My second fundamental objection to Prescott’s model is the absence of any independent corroborating evidence for the existence of what he calls technological shocks. This point is obviously crucial since Prescott treats technological shocks as the only driving force behind cyclical fluctuations. Prescott interprets all movements in measured total factor productivity as being the result of technology shocks or to a small extent measurement error. He provides no discussion of the source or nature of these shocks, nor does he cite any microeconomic evidence for their importance. I suspect that the vast majority of what Prescott labels technology shocks are in fact the observable concomitants of labor hoarding and other behavior which Prescott does not allow in his model. 24 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors SCEPTICAL OBSERVATIONS 99 Lawrence H.Summers Skeptical Observations Two observations support this judgment. First, it’s hard to find direct evidence of the existence of large technological shocks. Consider the oil shocks, certainly the most widely noted and commented on shocks of the postwar period. How much might they have been expected to reduce total factor productivity? In one of the most careful studies of this issue, Ernst Berndt (1980, p. 85) concludes that “energy price or quantity variations since 1973 do not appear to have had a significant direct role in the slowdown of aggregate labor productivity in U.S. manufacturing, 1973–77.” This is not to deny that energy shocks have important effects. But they have not accounted for large movements in measured total factor productivity. Prescott assumes that technological changes are irregular, but is unable to suggest any specific technological shocks which presage the downturns that have actually taken place. A reasonable challenge to his model is to ask how it accounts for the 1982 recession, the most serious downturn of the postwar period. More generally, it seems to me that the finding that measured productivity frequently declines is difficult to account for technologically. What are the sources of technical regress? Between 1973 and 1977, for example, both mining and construction displayed negative rates of productivity growth. For smaller sectors of the economy, negative productivity growth is commonly observed. A second observation casting doubt on Prescott’s assumed driving force is that while technological shocks leading to changes in total factor productivity are hard to find, other explanations are easy to support. Jon Fay and James Medoff (1985) surveyed some 170 firms on their response to downturns in the demand for their output. The questions asked were phrased to make clear that it was exogenous downturns in their output that were being inquired about. Fay and Medoff (1985, p. 653) summarize their results by stating that “the evidence indicates that a sizeable portion of the swings in productivity over the business cycle is, in fact, the result of firms’ decisions to hold labor in excess of regular production requirements and to hoard labor.” According to their data, the typical plant in the U.S. manufacturing sector paid for 8 percent more blue-collar hours than were needed for regular production work during the trough quarter of its most recent downturn. After taking account of the amount of other worthwhile work that was completed by blue-collar employees during the trough quarter, 4 percent of the blue-collar hours paid for were hoarded. Similar conclusions have been reached in every other examination of microeconomic data on productivity that I am aware of. In Prescott’s model, the central driving force behind cyclical fluctuations is technological shocks. The propagation mechanism is intertemporal substitution in employment. As I have argued so far, there is no independent evidence from any source for either of these phenomena. What About Prices?… My third fundamental objection to Prescott’s argument is that he does price-free economic analysis. Imagine an analyst confronting the market for ketchup. Suppose she or he decided to ignore data on the price of ketchup. This would considerably increase the analyst’s freedom in accounting for fluctuations in the quantity of ketchup purchased. Indeed, without looking at the price of ketchup, it would be impossible to distinguish supply shocks from demand shocks. It is difficult to believe that any explanation of fluctuations in ketchup sales that did not confront price data would be taken seriously, at least by hard-headed economists. Yet Prescott offers us an exercise in price-free economics. While real wages, interest rates, and returns to capital are central variables in his model, he never looks at any data on them except for his misconstrual of the average real interest rate over the postwar period. Others have confronted models like Prescott’s to data on prices with what I think can fairly be labeled dismal results. There is simply no evidence to support any of the price effects predicted by the model. Prescott’s work does not resolve—or even mention—the empirical reality emphasized by Robert Barro and Robert King (1982) that consumption and leisure move in opposite directions over the business cycle with no apparent procyclicality of real wages. It is finessed by ignoring wage data. Prescott’s own work with Rajnish Mehra (1985) indicates that the asset pricing implications of models like the one he considers here are decisively rejected by nearly 100 years of historical experience. I simply do not understand how an economic model can be said to have been tested without price data. I believe that the preceding arguments demonstrate that real business cycle models of the type surveyed by Prescott do not provide a convincing account of cyclical fluctuations. Even if this strong proposition is not accepted, they suggest that there is room for factors other than productivity shocks as causal elements in cyclical fluctuations. 25 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 100 FOUNDATIONS … And Exchange Failures? A fourth fundamental objection to Prescott’s work is that it ignores the fact that partial breakdowns in the exchange mechanism are almost surely dominant factors in cyclical fluctuations. Consider two examples. Between 1929 and 1933, the gross national product in the United States declined 50 percent, as employment fell sharply. In Europe today, employment has not risen since 1970 and unemployment has risen more than fivefold in many countries. I submit that it defies credulity to account for movements on this scale by pointing to intertemporal substitution and productivity shocks. All the more given that total factor productivity has increased more than twice as rapidly in Europe as in the United States. If some other force is responsible for the largest fluctuations that we observe, it seems quixotic methodologically to assume that it plays no role at all in other smaller fluctuations. Whatever mechanisms may have had something to do with the depression of the 1930s in the United States or the depression today in Europe presumably have at least some role in recent American cyclical fluctuations. What are those mechanisms? We do not yet know. But it seems clear that a central aspect of depressions, and probably economic fluctuations more generally, is a breakdown of the exchange mechanism. Read any account of life during the Great Depression in the United States. Firms had output they wanted to sell. Workers wanted to exchange their labor for it. But the exchanges did not take place. To say the situation was constrained Pareto optimal given the technological decline that took place between 1929 and 1933 is simply absurd, even though total factor productivity did fall. What happened was a failure of the exchange mechanism. This is something that no model, no matter how elaborate, of a long-lived Robinson Crusoe dealing with his changing world is going to confront. A model that embodies exchange is a minimum prerequisite for a serious theory of economic downturns. The traditional Keynesian approach is to postulate that the exchange mechanism fails because prices are in some sense rigid, so they do not attain market-clearing levels and thereby frustrate exchange. This is far from being a satisfactory story. Most plausible reasons why prices might not change also imply that agents should not continue to act along static demand and supply curves. But it hardly follows that ignoring exchange failures because we do not yet fully understand them is a plausible strategy. Where should one look for failures of the exchange process? Convincing evidence of the types of mechanisms that can lead to breakdowns of the exchange mechanism comes from analyses of breakdowns in credit markets. These seem to have played a crucial role in each of the postwar recessions. Indeed, while it is hard to account for postwar business cycle history by pointing to technological shocks, the account offered by, for example, Otto Eckstein and Allen Sinai (1986) of how each of the major recessions was caused by a credit crunch in an effort to control inflation seems compelling to me. Conclusion Even at this late date, economists are much better at analyzing the optimal response of a single economic agent to changing conditions than they are at analyzing the equilibria that will result when diverse agents interact. This unfortunate truth helps to explain why macroeconomics has found the task of controlling, predicting, or even explaining economic fluctuations so difficult. Improvement in the track record of macroeconomics will require the development of theories that can explain why exchange sometimes works well and other times breaks down. Nothing could be more counterproductive in this regard than a lengthy professional detour into the analysis of stochastic Robinson Crusoes. 26 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors SCEPTICAL OBSERVATIONS 101 Lawrence H.Summers Skeptical Observations References Altonji, Joseph G. 1982. The intertemporal substitution model of labour market fluctuations: An empirical analysis. Review of Economic Studies 49 (Special Issue): 783–824. Barro, Robert J., and King, Robert G. 1982. Time-separable preferences and intertemporal-substitution models of business cycles. Working Paper 888. National Bureau of Economic Research. Berndt, Ernst R. 1980. Energy price increases and the productivity slowdown in United States manufacturing. In The decline in productivity growth, pp. 60–89. Conference Series 22. Boston: Federal Reserve Bank of Boston. Eckstein, Otto, and Sinai, Allen. 1986. The mechanisms of the business cycle in the postwar era. In The American business cycle: Continuity and change, ed. Robert J.Gordon, pp. 39–105. National Bureau of Economic Research Studies in Business Cycles, vol. 25. Chicago: University of Chicago Press. Eichenbaum, Martin S.; Hansen, Lars P.; and Singleton, Kenneth J. 1986. A time series analysis of representative agent models of consumption and leisure choice under uncertainty. Working Paper 1981. National Bureau of Economic Research. Fay, Jon A., and Medoff, James L. 1985. Labor and output over the business cycle: Some direct evidence. American Economic Review 75 (September): 638–55. Ghez, Gilbert R., and Becker, Gary S. 1975. The allocation of time and goods over the life cycle. New York: National Bureau of Economic Research. Mankiw, N.Gregory; Rotemberg, Julio J.; and Summers, Lawrence H. 1985. Intertemporal substitution in macroeconomics. Quarterly Journal of Economics 100 (February): 225–51. Mehra, Rajnish, and Prescott, Edward C. 1985. The equity premium: A puzzle. Journal of Monetary Economics 15 (March): 145–61. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 102 CHAPTER 6 Federal Reserve Bank of Minneapolis Quarterly Review Fall 1986 Response to a Skeptic Edward C.Prescott Adviser Research Department Federal Reserve Bank of Minneapolis and Professor of Economics University of Minnesota New findings in science are always subject to skepticism and challenge. This is an important part of the scientific process. Only if new results successfully withstand the attacks do they become part of accepted scientific wisdom. Summers (in this issue) is within this tradition when he attacks the finding I describe (in this issue) that business cycles are precisely what economic theory predicts given the best measures of people’s willingness and ability to substitute consumption and leisure, both between and within time periods. I welcome this opportunity to respond to Summers’ challenges to the parameter values and the business cycle facts that I and other real business cycle analysts have used. In challenging the existing quality of measurement and not providing measurement inconsistent with existing theory, Summers has conceded the point that theory is ahead of business cycle measurement. Miscellaneous Misfires Before responding to Summers’ challenges to the measurements used in real business cycle analyses, I will respond briefly to his other attacks and, in the process, try to clarify some methodological issues in business cycle theory as well as in aggregate economic theory more generally. Prices Summers asks, Where are the prices? This question is puzzling. The mechanism real business cycle analysts use is the one he and other leading people in the field of aggregate public finance use: competitive equilibrium. Competitive equilibria have relative prices. As stated in the introduction of “Theory Ahead of Business Cycle Measurement” (in this issue), the business cycle puzzle is, Why are there large movements in the time allocated to market activities and little associated movements in the real wage, the price of people’s time? Along with that price, Kydland and I (1982, 1984) examine the rental price of capital. An infinity of other relative prices can be studied, but these are the ones needed to construct national income and product accounts. The behavior of these prices in our models conforms with that observed. In competitive theory, an economic environment is needed. For that, real business cycle analysts have used the neoclassical growth model. It is the preeminent model in aggregate economics. It was developed to account for the growth facts and has been widely used for predicting the aggregate effects of alternative tax schemes as well. With the labor/leisure decision endogenized, it is the appropriate model to study the aggregate implications of technology change uncertainty. Indeed, in 1977 Lucas, the person responsible for making business cycles again a central focus in economics, defined them (p. 23) as deviations from the neoclassical growth model—that is, fluctuations in hours allocated to market activity that are too large to be accounted for by changing marginal productivities of labor as reflected in real wages. Lucas, like me and virtually everyone else, assumed that, once characterized, the competitive equilibrium of the calibrated neoclassical growth economy would display much smaller fluctuations than do the actual U.S. data. Exploiting advances in theory and computational methods, Kydland and I (1982, 1984) and Hansen (1985) computed and studied the competitive equilibrium process for this model economy. We were surprised to find 28 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors RESPONSE TO A SCEPTIC 103 Edward C.Prescott Response to a Skeptic the predicted fluctuations roughly as large as those experienced by the U.S. economy since the Korean War. Some economists have been reluctant to use the competitive equilibrium mechanism to study business cycle fluctuations because they think it is contradicted by a real-world observation: some individuals who are not employed would gladly switch places with similarly skilled individuals who are. Solow (1986, p. S34), for example, predicted that “any interesting and useful solution to that riddle will almost certainly involve an equilibrium concept broader, or at least different from, pricemediated market-clearing.” Rogerson (1984) proved him wrong. If the world had no nonconvexities or moral hazard problems, Solow would be correct. But the mapping between time allocated to market activities and units of labor service produced does have nonconvexities. Time spent commuting is not producing labor services, yet it is time allocated to market activity. With nonconvexities, competitive equilibrium theory implies that the commodities traded or priced are complicated contracted arrangements which can include employment lotteries with both winners and losers. As shown by Hansen (1985), competitive theory accounts well for the observation that the principal margin of adjustment in aggregate hours is the number of people employed rather than the number of hours worked per person—as well as for the observation of so-called involuntary unemployment. Technology Shocks Another Summers question is, Where are the technology shocks? Apparently, he wants some identifiable shock to account for each of the half dozen postwar recessions. But our finding is not that infrequent large shocks produce fluctuations; it is, rather, that small shocks do, every period. At least since Slutzky (1927), some stable low-order linear stochastic difference equations have been known to generate cycles. They do not have a few large shocks; they have small shocks, one every period. The equilibrium allocation for the calibrated neoclassical growth model with persistent shocks to technology turns out to be just such a process. technology shocks as predicted by the neoclassical growth model. I do not argue that disruptions in the payment and credit system would not disrupt the economy. That theory predicts one factor has a particular nature and magnitude does not imply that theory predicts all other factors are zero. I only claim that technology shocks account for more than half the fluctuations in the postwar period, with a best point estimate near 75 percent. This does not imply that public finance disturbances, random changes in the terms of trade, and shocks to the technology of exchange had no effect in that period. Neither do I claim that theory is ahead of macroeconomic measurement in all respects. As Summers points out, Mehra and I (1985) have used the representative agent construct to predict the magnitude of the average risk premium of an equity claim over a real bill. Our predicted quantity is small compared to the historically observed average difference between the yields of the stock market and U.S. Treasury bills. But this is not a failure of the representative agent construct; it is a success. We used theory to predict the magnitude of the average risk premium. That the representative agent model is poorly designed to predict differences in borrowing and lending rates—to explain, for example, why the government can borrow at a rate at least a few percentage points less than the one at which most of us can borrow—does not imply that this model is not well designed for other purposes—for predicting the consequences of technology shocks for fluctuations at the business cycle frequencies, for example. Measurement Issues Summers challenges the values real business cycle analysts have selected for three model parameters. By arguing that historically the real U.S. interest rate is closer to 1 percent than to the model economy’s approximately 4 percent, he is questioning the value selected for the subjective time discount factor. He explicitly questions our value for the leisure share parameter. And Summers’ challenge to the observation that labor productivity is procyclical is implicitly a challenge to my measure of the technology shock variance parameter. My Claims Real Interest Rate Summers has perhaps misread some of my review of real business cycle research (in this issue). There I do not argue that the Great American Depression was the equilibrium response to Summers points out that the real return on U.S. Treasury bills over the last 30 years has been about 1 percent, which is far from the average real interest rate of the economies that Kydland 29 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 104 FOUNDATIONS and I have studied. But for the neoclassical growth model, the relevant return is not the return on T-bills. It is the return on tangible capital, such things as houses, factories, machines, inventories, automobiles, and roads. The return on capital in the U.S. business sector is easily calculated from the U.S. National Income and Product Accounts, so we use it as a proxy for the return on U.S. capital more generally. This number is obtained by dividing the income of capital net of the adjusted capital consumption allowance by the capital stock in the business sector. For the postwar years, the result is approximately 4 percent, about the average real return for the model economies. Preferences Summers also questions the value of the leisure share parameter and argues that it is not well tied down by micro observation at the household level, as we claim. This is a potentially important parameter. If it is large, the response of labor supply to temporary changes in the real wage is large. Only if that response is large will large movements in employment be associated with small co-movements in the real wage. Kydland and I conclude that the leisure share parameter is not large based on findings reported by Ghez and Becker (1975). They report (p. 95) that the annual productive time endowment of U.S. males is 5,096 hours. They also say (p. 95) that U.S. females allocate about 75 hours per week to personal care, leaving 93 hours of production time per week. This multiplied by 52 is 4,836 hours, the annual productive time endowment of females. Ghez and Becker also report the average annual hours of employment for noninstitutionalized, working-age males as about 2,000 hours (pp. 85–91). If females allocate half as many hours to market employment as do males, the average fraction of time the U.S. working-age population spends in employment is about 0.30. Adding to this the time spent commuting yields a number close to those for our models. (They are all between 0.30 and 0.31 in Kydland and Prescott 1982 and 1984.) Initially Kydland and I used time additive preferences, and the predictions of theory for productivity movements were as large in percentage terms as aggregate hour movements. This is inconsistent with observations, so I did not take seriously the prediction of theory that a little over half the aggregate output fluctuations in the postwar period were responses to technology shocks. At that time, measurement was still ahead of theory. Then, the prediction of theory would have been consistent with the relative movement of productivity and aggregate hours, and technology shocks would have accounted for the business cycle phenomena, if the leisure share parameter were five-sixths. With the discipline we used, however, this share parameter had to be consistent with observations on household time allocation. That we are now debating about a theory of aggregate phenomena by focusing on household time allocation is evidence that economic theory has advanced. Now, like physical scientists, when economists model aggregate phenomena, the parameters used can be measured independently of those phenomena. In our 1982 paper, Kydland and I did claim that fluctuations of the magnitude observed could plausibly be accounted for by the randomness in the technological change process. There we explored the implications of a distributed lag of leisure being an argument of the period utility function rather than just the current level of leisure. Like increasing the leisure share parameter, this broadening results in larger fluctuations in hours in response to technology shocks. Kydland (1983) then showed that an unmeasured household-specific capital stock could rationalize this distributed lag. In addition, the lag was not inconsistent with good micro measurement, and these parameters could be measured independently of the business cycle phenomena. The distributed lag was a long shot, though, so we did not claim that theory had caught up to measurement. Since then, however, two panel studies found evidence for a distributed lag of the type we considered (Hotz, Kydland, and Sedlacek 1985; Eckstein and Wolpin 1986). With this development, theory and measurement of the business cycle became roughly equal. Subsequently, an important advance in aggregate theory has made moot the issue of whether Kydland’s and my assumed preferences for leisure are supported by micro measurement. Given an important nonconvexity in the mapping between time allocated to market activities and units of labor service produced, Rogerson (1984) showed that the aggregate elasticity of labor supply to temporary changes in the real wage is large independent of individuals’ willingness to intertemporally substitute leisure. This nicely rationalized the disparate micro and macro labor findings for this elasticity—the microeconomists’ that it is small (for example, Ashenfelter 1984) and the macroeconomists’ that it is large (for example, Eichenbaum, Hansen, and Singleton 1984). Hansen (1985) 30 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors RESPONSE TO A SCEPTIC 105 Edward C.Prescott Response to a Skeptic introduced this nonconvexity into the neoclassical growth model. He found that with this feature theory predicts that the economy will display the business cycle phenomena even if individuals’ elasticity of labor supply to temporary changes in the real wage is small. Further, with this feature he found theory correctly predicts that most of the variation in aggregate hours of employment is accounted for by variation in the number of people employed rather than in the number of hours worked per person. Technology Uncertainty In our 1982 paper, Kydland and I searched over processes for the technological change process. We did sensitivity analysis with the other parameters, but found the conclusions relatively insensitive to their assumed values (except for the distributed lag of leisure parameters just discussed). The parameters of the technological change process did affect our predictions of the aggregate implications of uncertainty in the technology parameter. In fact, Lucas (1985, p. 48) criticized us for searching for the best fit. In “Theory Ahead of Business Cycle Measurement,” I directly examined the statistical properties of the technology coefficient process. I found that the process is an approximate random walk with standard deviation of change in the logs approximately 0.00763 per quarter. When this number is used in the Hansen model, fluctuations predicted are even larger than those observed. In Kydland’s and my model (1984), they are essentially equal to those observed. Some, on the basis of theory, think that the factors producing technological change are small, many, and roughly uncorrelated. If so, by the law of large numbers, these factors should average out and the technological change process should be very smooth. I found (in this issue) empirical evidence to the contrary. Others have too. Summers and Heston (1984) report the annual gross national products for virtually every country in the postwar period. They show huge variation across countries in the rate of growth of per capita income over periods sufficiently long that business cycle variations are a minor consideration. Even individual countries have large variation in the decade growth rates of per capita output. Given Solow’s (1957) finding that more than 75 percent of the changes in per capita output are accounted for by changes in the technology parameter, the evidence for variation in the rate of technological advance is strong. Obviously, economists do not have a good theory of the determinants of technological change. In this regard, measurement is ahead of theory. The determinants of the rate of technological change must depend greatly on the institutions and arrangements that societies adopt. Why else should technology advance more rapidly in one country than in another or, within a country, more rapidly in one period than in another? But a theory of technological change is not needed to predict responses to technological change. The key parameter is the variance of the technology shock. This is where better measurement could alter the prediction of theory. Is measuring this variance with Solow’s (1957) method (as I did) reasonable? I showed that measures of the technology shock variance are insensitive to cyclical variations in the capital utilization rate. Even if that rate varies proportionately to hours of employment and the proportionality constant is selected so as to minimize the measured standard deviation of the technology shock, that measured deviation is reduced only from 0.00763 to 0.00759. Further, when the capital utilization rate varies in this way for the model, the equilibrium responses are significantly larger. Variation in the capital utilization rate does not appear to greatly bias my estimate of the importance of technological change variance for aggregate fluctuations. Perhaps better measurement will find that the technological change process varies less than I estimated. If so, a prediction of theory is that the amount of fluctuation accounted for by uncertainty in that process is smaller. If this were to happen, I would be surprised. I can think of no plausible source of measurement error that would produce a random walk-like process for technological change. Labor Hoarding Summers seems to argue that measured productivity is procyclical because measurement errors are cyclical. To support his argument, he cites a survey by Fay and Medoff (1985), which actually has little if anything to say about cyclical movements. Fay and Medoff surveyed more than 1,000 plant managers and received 168 usable responses. One of the questions asked was, How many extra blue-collar workers did you have in your recent downturn? They did not ask, How many extra workers did you have at the trough quarter and at the peak quarter of the most recent business cycle? Answers to those questions are 30 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 106 FOUNDATIONS needed to conclude how the number of extra blue-collar workers reported by managers varies over the cycle. Even if these questions had been asked, though, the response to them would not be a good measure of the number of redundant workers. Such questions are simply too ambiguous for most respondents to interpret them the same way. The argument that labor hoarding is cyclical is not supported by theory either. The fact that labor is a quasi-fixed factor of production in the sense of Oi (1962) does not imply that more workers will be hoarded in recessions than in expansions. In bad times a firm with low output may be less reluctant to lay off workers than in good times because the worker is less likely to be hired by another firm. This argument suggests that labor hoarding associated with firm-specific output variations should be procyclical. Leisure consumed on the job also may be less in bad times than in good because work discipline may be greater. That is, an entrepreneur might be less reluctant to fire a worker in bad times because the worker can more easily be replaced. One might reasonably think, therefore, that labor’s quasi-fixed nature makes measured productivity less, not more, cyclically volatile than productivity really is. There is another, better reason to think that. In the standard measures of aggregate hours of employment, the hours of an experienced MBA from one of the best business schools are treated the same as those of a high school dropout. Yet these hours do not on average command the same price in the market, which is evidence that they are not the same commodity. In the neoclassical growth model, the appropriate way to aggregate hours is in terms of effective units of labor. That is, if the MBA’s productivity is five times that of the high school dropout, then each hour of the MBA’s time is effectively equivalent to five hours of the high school dropout’s time. The work of Kydland (1984) suggests this correction is an important one. The more educated and on average more highly paid have much less variability in annual hours of employment than do the less educated. Kydland (1984, p. 179) reports average hours and average wages as well as sensitivity of hours to the aggregate unemployment rate for adult males categorized by years of schooling. His figures imply that a 1 percentage point change in the aggregate unemployment rate for adult males is associated with a 1.24 percent change in equally weighted hours. When those hours are measured as effective units of labor, the latter change is only 0.65 percent. This is strong evidence that if the labor input were measured correctly, the measure of productivity would vary more. To summarize, measurement of the labor input needs to be improved. By questioning the standard measures, Summers is agreeing that theory is ahead of business cycle measurement. More quantitative theoretic work is also needed, to determine whether abstracting from the fact that labor is a partially fixed factor affects any of the real business cycle models’ findings. Of course, introducing this feature—or others—into these models may significantly alter their predictions of the aggregate implications of technology uncertainty. But respectable economic intuition must be based on models that have been rigorously analyzed. To Conclude Summers cannot be attacking the use of competitive theory and the neoclassical growth environment in general. He uses this standard model to predict the effects of alternative tax policies on aggregate economic behavior. He does not provide criteria for deciding when implications of this model should be taken seriously and when they should not be. My guess is that the reason for skepticism is not the methods used, but rather the unexpected nature of the findings. We agree that labor input is not that precisely measured, so neither is technological uncertainty. In other words, we agree that theory is ahead of business cycle measurement. 31 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors RESPONSE TO A SCEPTIC 107 Edward C.Prescott Response to a Skeptic References Ashenfelter, O. 1984. Macroeconomic analyses and microeconomic analyses of labor supply. In Essays on macroeconomic Implications of financial and labor markets and political processes, ed. K.Brunner and A.H.Meltzer. Carnegie-Rochester Conference Series on Public Policy 21:117–55. Amsterdam: North-Holland. Eckstein, Zvi, and Wolpin, Kenneth I. 1986. Dynamic labor force participation of married women and endogenous work experience. Manuscript. Tel Aviv University and Ohio State University. Eichenbaum, Martin S.; Hansen, Lars P.; and Singleton, Kenneth J. 1984. A time series analysis of representative agent models of consumption and leisure choice under uncertainty. Working paper, CarnegieMellon University. Fay, Jon A., and Medoff, James L. 1985. Labor and output over the business cycle: Some direct evidence. American Economic Review 75 (September): 638–55. Ghez, Gilbert R., and Becker, Gary S. 1975. The allocation of time and goods over the life cycle. New York: National Bureau of Economic Research. Hansen, Gary D. 1985. Indivisible labor and the business cycle. Journal of Monetary Economics 16 (November): 309–27. Hotz, V.S.; Kydland, F.E.; and Sedlacek, G.L. 1985. Intertemporal preferences and labor supply. Working paper, Carnegie-Mellon University. Kydland, Finn E. 1983. Nonseparable utility and labor supply. Working paper, Hoover Institution. ____. 1984. Labor-force heterogeneity and the business cycle. In Essays on macroeconomic implications of financial and labor markets and political processes, ed. K.Brunner and A.H.Meltzer. Carnegie-Rochester Conference Series on Public Policy 21:173– 208. Amsterdam: North-Holland. Kydland, Finn E., and Prescott, Edward C. 1982. Time to build and aggregate fluctuations. Econometrica 50 (January): 1345–70. ____. 1984. The workweek of capital and labor. Research Department Working Paper 267. Federal Reserve Bank of Minneapolis. Lucas, Robert E., Jr. 1977. Understanding business cycles. In Stabilization of the domestic and international economy, ed. K.Brunner and A.H.Meltzer. Carnegie-Rochester Conference Series on Public Policy 5:7– 29. Amsterdam: North-Holland. ____. 1985. Models of business cycles. Manuscript prepared for the Yrjo Jahnsson Lectures , Helsinki, Finland. University of Chicago. Mehra, Rajnish, and Prescott, Edward C. 1985. The equity premium: A puzzle. Journal of Monetary Economics 15 (March): 145–61. Oi, Walter Y. 1962. Labor as a quasi-fixed factor. Journal of Political Economy 70 (December): 538–55. Rogerson, R.D. 1984. Indivisible labor, lotteries and equilibrium. Economics Department Working Paper 10. University of Rochester. Slutzky, Eugen. 1927. The summation of random causes as the source of cyclic processes. In Problems of economic conditions, ed. Conjuncture Institute, Moskva (Moscow), vol. 3, no. 1. Revised English version, 1937, in Econometrica 5:105–46. Solow, Robert M. 1957. Technical change and the aggregate production function. Review of Economics and Statistics 39 (August): 312– 20. _____. 1986. Unemployment: Getting the questions right. Economica 53 (Supplement): S23–S34. Summers, Robert, and Heston, Alan. 1984. Improved international comparisons of real product and its composition: 1950–1980. Review of Income and Wealth 30 (June): 207–62. 33 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 108 CHAPTER 7 Journal of Monetary Economics 21 (1988) 195–232. North-Holland PRODUCTION, GROWTH AND BUSINESS CYCLES I. The Basic Neoclassical Model Robert G.KING, Charles I.PLOSSER and Sergio T.REBELO* University of Rochester, Rochester, NY 14627, USA Received September 1987, final version received December 1987 This paper presents the neoclassical model of capital accumulation augmented by choice of labor supply as the basic framework of modern real business cycle analysis. Preferences and production possibilities are restricted so that the economy displays steady state growth. Then we explore the implications of the basic model for perfect foresight capital accumulation and for economic fluctuations initiated by impulses to technology. We argue that the neoclassical approach holds considerable promise for enhancing our understanding of fluctuations. Nevertheless, the basic model does have some important shortcomings. In particular, substantial persistence in technology shocks is required if the model economy is to exhibit periods of economic activity that persistently deviate from a deterministic trend. 1. Introduction and summary Real business cycle analysis investigates the role of neoclassical factors in shaping the character of economic fluctuations. In this pair of essays, we provide an introduction to the real business cycle research program by considering the basic concepts, analytical methods and open questions on the frontier of research. The focus of the present essay is on the dynamic aspects of the basic neoclassical model of capital accumulation. This model is most frequently encountered in analyses of economic growth, but we share Hicks’ (1965, p. 4) perspective that it is also a basic laboratory for investigating more general dynamic phenomena involving the choice of consumption, work effort and investment. Our use of the neoclassical model of capital accumulation as the engine of analysis for the investigation of economic fluctuations raises a number of central issues. First, what role does economic growth play in the study of * The authors acknowledge financial support from the National Science Foundation. King and Plosser have joint affiliations with the Department of Economics and the W.E.Simon Graduate School of Business, University of Rochester. Rebelo is affiliated with the Department of Economics, University of Rochester and the Department of Economics, Portuguese Catholic University. We have benefited from the comments of Andrew Abel and Larry Christiano, as well as from those of seminar participants at the Federal Reserve Bank of Richmond, Brasenose College, Oxford, Institute for International Economic Studies, University of Stockholm, Northwestern University, Yale University, and Columbia University. 0304–3932/88/$3.50 ©1988, Elsevier Science Publishers B.V. (North-Holland) © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE BASIC NEOCLASSIC MODEL 196 109 R.G.King et al., Production, growth and business cycles I economic fluctuations? More precisely, does the presence of economic growth restrict the preference and production specifications in ways that are important for the analysis of business cycles? Second, what analytical methods can be employed to study the time series implications of the neoclassical model? Third, what are the dynamics of the neoclassical model in response to technology shocks? Finally, does the neoclassical model—driven by technology shocks—replicate important features of macroeconomic time series? The analysis of these issues forms the core of the present paper and establishes the building blocks of real business cycle theory. Real business cycle theory, though still in the early stages of development, holds considerable promise for enhancing our understanding of economic fluctuations and growth as well as their interaction. The basic framework developed in this essay is capable of addressing a wide variety of issues that are commonly thought to be important for understanding business cycles. While we focus here on models whose impulses are technological, the methods can be adapted to consider shocks originating from preferences or other exogenous factors such as government policies and terms of trade. Some of these extensions to the basic framework are developed in the companion essay. To many readers it must seem heretical to discuss business cycles without mentioning money. Our view, however, is simply that the role of money in an equilibrium theory of economic growth and fluctuations remains an open area for research. Further, real disturbances generate rich and neglected interactions in the basic neoclassical model that may account for a substantial portion of observed fluctuations. The objective of real business cycle research is to obtain a better understanding of the character of these real fluctuations. Without an understanding of these real fluctuations it is difficult a priori to assign an important role to money. The organization of the paper follows the sequence of questions outlined above. We begin in section 2 by describing the preferences, endowments and technology of the basic (one-sector) neoclassical model of capital accumulation.1 In contrast to the familiar textbook presentation of this model, however, work effort is viewed as a choice variable. We then discuss the restrictions on production possibilities and preferences that are necessary for steady state growth. On the production side, with a constant returns to scale production function, technical progress must be expressible in labor augmenting (Harrod neutral) form. In a feasible steady state, it follows that consumption, investment, output and capital all must grow at the exogenously specified rate of technical change. On the other hand, since the endowment of time is constant, work effort cannot grow in the steady state. Thus, preferences must be restricted so that there is no change in the level of effort on the steady state growth path despite the rise in marginal productivity stemming from technical progress, 1 A more detailed and unified development of the material is presented in the technical appendix, available from the authors on request. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 110 FOUNDATIONS R.G.King et al., Production, growth and business cycles I 197 i.e., there must be an exact offset of appropriately defined income and substitution effects. Section 3 concerns perfect foresight dynamic competitive equilibria, which we analyze using approximations near the steady state. Using a parametric version of the model, with parameters chosen to match the long-run U.S. growth experience, we study the interaction between intertemporal production possibilities and the equilibrium quantity of labor effort. Off the steady state path, we find that capital and effort are negatively related despite the fact that the marginal product of labor schedule is positively related to the capital stock. That is, in response to the high real rate of return implied by a low capital stock, individuals will substitute intertemporally to produce additional resources for investment. Working from a certainty equivalence perspective, section 4 considers how temporary productivity shocks influence economic activity, generating ‘real business cycles’ in the terminology of Long and Plosser (1983). Again there is an important interaction between variation in labor input—this time in response to a productivity shock—and the intertemporal substitution in production permitted by capital accumulation. Purely temporary technology shocks call forth an expansion of labor input once the Long and Plosser (1983) assumption of complete depreciation is replaced by a more realistic value,2 since more durable capital increases the feasibility of intertemporal substitution of goods and leisure. Nevertheless, with purely temporary productivity shocks, we find that there are important deficiencies of the basic neoclassical model. Although there is substantial serial correlation in consumption and capital as a consequence of consumption smoothing, there is effectively no serial correlation in output or employment. This lack of propagation reflects two basic properties of the parameterized model: (i) a negative relation between capital and effort along the transition path and (ii) the minor effect of a purely temporary technology shock on a large and durable capital stock. Thus, the basic neoclassical capital accumulation mechanism is important for permitting intertemporal substitution of goods and leisure, but it does not generate serial correlation in output and employment close to that exhibited by macroeconomic data. It is necessary, therefore, to incorporate substantial serial correlation in productivity shocks [as in Kydland and Prescott (1982), Long and Plosser (1983), Hansen (1985), and Prescott (1986)] if the basic neoclassical model is to generate business fluctuations that resemble those in post-war U.S. experience. Since serial correlation involves movements in productive opportunities that are more persistent in character, labor input responds less elastically to a given size shock, but its response remains positive. On the other hand, with more persistent productivity shocks, consumption responds more elastically in accord with the permanent income theory. 2 By a purely temporary shock, we mean one that lasts for a single time period, which is taken to be a quarter in our analysis. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE BASIC NEOCLASSIC MODEL 198 111 R.G.King et al., Production, growth and business cycles I In section 5, we show that the basic neoclassical model—with persistent technology shocks—captures some key features of U.S. business cycles. For example, the model replicates observed differences in volatility across key series. Measured as a percentage of the standard deviation of output, there is an identical ordering of the model’s implications for investment, wages, consumption and hours, and the U.S. time series: investment is most volatile, followed by wages, consumption and then hours. But there are also aspects of the data that are poorly captured by the single-shock model. For example, consumption, investment and hours are much more highly correlated with output in the model than in the data. Professional interest in real business cycle analysis has been enhanced by the comparison of moments implied by neoclassical models with those of U.S. time series, as initiated by Kydland and Prescott (1982). Our implications for moments differ from those of Hansen (1985) and Prescott (1986), principally because we do not filter actual and model-generated time series to remove slow-moving components. For example, in Hansen’s and Prescott’s analyses, filtered hours and output have virtually identical volatilities, in both the model and the transformed data. By contrast, in our analysis, the volatility of hours is about half that of output (both in our model and post-war detrended U.S. data). These differences occur despite the fact that there is little economic difference in the models under study. Section 6 provides a brief summary and some concluding remarks. 2. The basic neoclassical model Our analysis of economic growth and fluctuations starts by summarizing the key features of the basic one-sector, neoclassical model of capital accumulation. Much of the discussion in this section will be familiar to readers of Solow (1956), Cass (1965), Koopmans (1965) and subsequent textbook presentations of their work, but it is important to build a base for subsequent developments. 2.1. Economic environment We begin by considering the preferences, technology and endowments of the environment under study. Preferences. We consider an economy populated by many identical infinitelylived individuals with preferences over goods and leisure represented by (2.1) where Ct is commodity consumption in period t and Lt is leisure in period © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 112 FOUNDATIONS R.G.King et al., Production, growth and business cycles I 199 t. Consumption and leisure are assumed throughout to be goods, so that utility is increasing in Ct and Lt.3 Production possibilities. There is only one final good in this economy and it is produced according to a constant returns to scale neoclassical production technology given by (2.2) where Kt is the predetermined capital stock (chosen at t-1) and Nt is the labor input in period t.4 We permit temporary changes in total factor productivity through At. Permanent technological variations are restricted to be in labor productivity, Xt, for reasons that we discuss below. Capital accumulation. In this simple neoclassical framework the commodity can be either consumed or invested. The capital stock evolves according to (2.3) where It is gross investment and δK is the rate of depreciation of capital.5 Resource constraints. In each period, an individual faces two resource constraints: (i) total time allocated to work and leisure must not exceed the endowment, which is normalized to one, and (ii) total uses of the commodity must not exceed output. These conditions are (2.4) (2.5) Naturally, there are also the non-negativity constraints Lt≥0, Nt≥0, Ct≥0 and Kt≥0. 3 Momentary utility, u(·), is assumed to be strictly concave and twice continuously differentiable. Further, it satisfies the Inada conditions, namely that limc→0D1u(c, L)=∞ and limc→∞D1u(c, L)=0, limL→0D2u(c, L)=∞ and limL→1D2u(c, L)=0, where Diu(·) is the first partial derivative of u(·) with respect to the function’s i th argument. 4 By neoclassical, we mean that the production function is concave, twice continuously differentiable, satisfies the Inada conditions, and that both factors are essential in production. 5 We abstract from adjustment costs to capital accumulation throughout the analysis, as these seem to us to be basically a restricted formulation of the two-sector neoclassical model. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE BASIC NEOCLASSIC MODEL 200 113 R.G.King et al., Production, growth and business cycles I 2.2. Individual optimization and competitive equilibrium The standard neoclassical analysis focuses on the optimal quantities chosen by a ‘social planner’ or representative agent directly operating the technology of the economy. Since our setup satisfies the conditions under which the second welfare theorem is valid, optimal capital accumulation will also be realized in a competitive equilibrium.6 In the companion essay, we discuss departures from the strict representative agent model including government expenditures and distorting taxes, productive externalities, and heterogeneity of preferences and productivities. In these contexts, we will need to be more precise about distinguishing between individual choices and competitive outcomes. 2.3. Steady state growth A characteristic of most industrialized economies is that variables like output per capita and consumption per capita exhibit sustained growth over long periods of time. This long-run growth occurs at rates that are roughly constant over time within economies but differ across economies. We interpret this pattern as evidence of steady state growth, by which we mean that the levels of certain key variables grow at constant—but possibly different—rates, at least some of which are positive. Additional restrictions on preferences and technologies are required if the system is to exhibit steady state growth. Restrictions on production. For a steady state to be feasible, Swan (1963) and Phelps (1966) show that permanent technical change must be expressible in a labor augmenting form, which rationalizes our specification in (2.2) above. To make for an easier comparison with other studies, we adopt the Cobb-Douglas production process for the bulk of our analysis, (2.6) where the quantity NtXt is usually referred to as effective labor units.7 Since variation in At is assumed temporary, we can ignore it for our investigation of steady state growth. The production function (2.6) and the accumulation equation (2.3) then imply that the steady state rates of growth of output, consumption, capital and investment per capita are all equal to the growth rate of labor augmenting technical progress.8 Denoting 6 The basic reference is Debreu (1954). See also Prescott and Lucas (1972). We note, however, that if technological change is labor augmenting, then the observed invariance of factor shares to the scale of economic activity cannot be used to rationalize the restriction to the Cobb-Douglas form. In the presence of labor augmenting technological progress the factor shares are constant for any constant returns to scale production function. 7 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 114 FOUNDATIONS R.G.King et al., Production, growth and business cycles I 201 one plus the growth rate of a variable Z as γz (i.e., Zt+1/Zt), then any feasible steady state requires (2.7a) and the growth rate of work effort to be zero, i.e., (2.7b) Since time devoted to work N is bounded by the endowment, it cannot grow in the steady state (2.7b). Thus, the only admissible constant growth rate for N is zero. In any such feasible steady state, the marginal product of capital and the marginal product of a unit of labor input in efficiency units are constant. The levels of the marginal products, however, depend on the ratio of capital to effective labor, which is not determined by the restriction to a feasible steady state. Restrictions on preferences. Eqs. (2.7a) and (2.7b) describe the technologically feasible steady state growth rates. If these conditions are not compatible with the efficiency conditions of agents in the economy, then they are of little interest since they would never be an equilibrium outcome. We can insure that the feasible steady state is compatible with an (optimal) competitive equilibrium, however, by imposing two restrictions on preferences: (i) the intertemporal elasticity of substitution in consumption must be invariant to the scale of consumption and (ii) the income and substitution effects associated with sustained growth in labor productivity must not alter labor supply. The first condition must hold because the marginal product of capital, which equals one plus the real interest rate in equilibrium, must be constant in the steady state. Since consumption is growing at a constant rate and the ratio of discounted marginal utilities must equal one plus the interest rate, the intertemporal elasticity of substitution must be constant and independent of the level of consumption. The second condition is required because hours worked cannot grow (γN=1) in the steady state. To reconcile this with a growing marginal productivity of labor—induced by labor augmenting technical change (Xt)— income and substitution effects of productivity growth must have exactly offsetting effects on labor supply (N).9 8 9 This result, in fact, holds for any constant returns to scale production function. Effective labor (NXt) will continue to grow at rate γx. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE BASIC NEOCLASSIC MODEL 8 202 115 R.G.King et al., Production, growth and business cycles I These conditions imply the following class of admissible utility functions:10 (2.8a) for 0<<1 and >1, while for =1, (2.8b) Some additional restrictions are necessary to assure that (i) consumption and leisure are goods and (ii) that utility is concave.11 The constant intertemporal elasticity of substitution in consumption is 1/ for these utility functions. For the remainder of our analysis we restrict ourselves to utility functions of this class. The requirement that preferences be compatible with steady state growth has important implications for the study of economic fluctuations. If there is no capital [i.e., if the production function is just of the form At(NtXt)], then there will be no response of hours to variation in Xt or At in general equilibrium. This arises because (i) utility implies that the income and substitution effects of wage changes just offset and (ii) with no intertemporal substitution in production, income effects must be fully absorbed within any decision period [as in Barro and King (1984)]. Thus, in all of the parameterizations of the neoclassical model that we consider, variations in work effort are associated with intertemporal substitution made possible in equilibrium by capital accumulation. 2.4. Stationary economies and steady states The standard method of analyzing models with steady state growth is to transform the economy into a stationary one where the dynamics are more amenable to analysis. In the context of the basic neoclassical model, this transformation involves dividing all variables in the system by the growth component X, so that c=C/X, k=K/X, i=I/X, etc. This economy is identical to a simple ‘no-growth’ economy with two exceptions. First the capital accumulation equation, Kt+1=(1–δK)Kt+It, becomes γXkt+1=(1-δK)kt+it. Second, transforming consumption in the preference specification generally alters the effective rate of time preference. That is, 10 See the technical appendix for a demonstration of the necessity of these conditions and that they imply (2.8a) and (2.8b). 11 When momentary utility is additively separable (2.8b), all that we require is that (L) is increasing and concave. When momentary utility is multiplicatively separable, then we require that (L) is (i) increasing and concave if <1 and (ii) decreasing and convex if >1. Defining Dn(L) as the n th total derivative of the function (L), we further require that— [LD2(L)/D(L)]>(1-)[LD(L)/(L)] to assure overall concavity of u(·). © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 116 FOUNDATIONS R.G.King et al., Production, growth and business cycles I 203 (2.9a) (2.9b) where β*=β(γx)1-σ and β*<1 is required throughout to guarantee finiteness of lifetime utility. Thus, unless =1, β*≠β. By suitable selection of X0, we Combining the can in either case make the objective resource constraints, we form the Lagrangian (2.10) The efficiency conditions for the transformed economy are (2.11)– (2.15). In these expressions, Di is the first partial derivative operator with respect to the ith argument. For convenience, we discount the Lagrange multipliers, i.e. (2.11) (2.12) (2.13) (2.14) (2.15) where (2.11)–(2.14) must hold for all t=1, 2, …∞ and (2.15) is the so-called transversality condition. The economy’s initial capital stock, k0, is given. Optimal per capita quantities for this economy—for a given sequence of technology shifts—are sequences of consumption effort capital stock and shadow prices that satisfy the efficiency conditions (2.11)–(2.15). Under our assumptions about © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE BASIC NEOCLASSIC MODEL 204 117 R.G.King et al., Production, growth and business cycles I preferences and production possibilities, conditions (2.11)–(2.15) are necessary and sufficient for an optimum.12 The prices that decentralize the optimal solution as a competitive equilibrium and the optimal sequences can be computed using the technology shifts and For instance, in a complete initial date markets framework the sequence of equilibrium prices of labor and the final good are, and Under perfect foresight respectively, (rational expectations), a regime of sequential loan markets and spot markets in labor services also supports the optimal solution as a competitive equilibrium. In this market structure, the relevant prices are the real interest rate between t and t+1, rt, and the real wage rate, wt. It is easy to demonstrate that these are given by (1+rt)=γXλt/λt+1β* and wt=AtD2F(kt, Nt). 3. Perfect foresight capital accumulation A major feature of the basic one sector neoclassical model with stationary technology is that the optimal capital stock converges monotonically to a stationary point.13 While qualitative results such as the monotonicity property are important, we wish to undertake quantitative analyses of capital stock dynamics. This requires that we exploit the fact that (2.11)– (2.14) can be reduced to a non-linear system of first-order difference equations in k and λ or a second-order equation in k only. The two boundary conditions of this system are the transversality condition (2.15) and the initial capital stock, k0. We focus on approximate linear dynamics in the neighborhood of the steady state denoted by (A, k, N, c and y).14 3.1. Approximation method The initial step in obtaining the system of linear difference equations is to approximate (2.11)–(2.14) near the stationary point. To do this, we express each condition in terms of the percentage deviation from the stationary value, which we indicate using a circumflex Then, we linearize each condition 12 See Weitzman (1973) and Romer and Shinotsuka (1987). In the fixed labor case, which is the most thoroughly studied, this property has been shown to derive principally from preferences, in that the concavity of u(·) is sufficient for monotonicity so long as there is a maximum sustainable capital stock [Boyd (1986) and Becker et al. (1986)]. In environments such as ours, where the production function is strictly concave in capital (for fixed labor), monotonicity also insures that capital approaches a unique stationary point. 14 The technical appendix discusses solution methods in considerable detail. The linear approximation method, it should be noted, rules out certain phenomena that may arise in the basic neoclassical model, such as a humped shaped transition path for investment [see King (1987)]. 13 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 118 FOUNDATIONS R.G.King et al., Production, growth and business cycles I 205 in terms of deviations from the stationary point. The results for the first two conditions can be written as follows: (3.1) (3.2) where ξab is the elasticity of the marginal utility of a with respect to b.15 Approximation of the intertemporal efficiency condition (2.13) implies that (3.3) where ηA is the elasticity of the gross marginal product of capital with respect to A evaluated at the steady state, etc.16 Finally, approximation of the resource constraint (2.14) implies (3.4) where sc and si are consumption and investment shares in output and As in other linear optimal control settings, expressions (3.1) and (3.2) as functions of the state can be solved to give optimal decisions and the co-state (shadow price) Further, given these variables (conditionally) optimal decisions, expressions (3.3) and (3.4) imply a firstorder dynamic system in and (3.5) where W is a 2×2 matrix and R and Q are 2×1 vectors. To compute the solution to this difference equation and to examine its properties, we use the decomposition W=PµP –1, where P is the matrix of characteristic vectors of W and µ is a diagonal matrix with the characteristic roots on the diagonal. Ordering the roots (µ1, µ2) in increasing absolute value, it can be shown that 15 When the utility function is additively separable, it follows that ξ cc=–1, ξ cl=ξ lc=0 and ξ ll=LD2(L)/D(L). When the utility function is multiplicatively separable, it follows that ξ cc=-, ξ cl=LD(L)/v(L), ξ lc=1- and ξ ll =LD2(L)/D(L). 16 With the Cobb-Douglas assumption, it follows that ηA=[γ X-*(1-δK)]/γX, ηk=αηA and ηN=αηA. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE BASIC NEOCLASSIC MODEL 206 119 R.G.King et al., Production, growth and business cycles I 0<µ1<1<β*–1<µ2. The general solution to the difference equation for and is given by specified initial conditions (3.6) Since Wt=PµtP–1 and the root µ2 exceeds (β *)–1>1, it follows that the system is on an explosive path and thus violates the transversality condition for There is a specific value of the initial shadow price however, arbitrary that results in (3.6) satisfying the transversality condition (2.15). This particular solution specifies the unique optimal (and competitive and shadow prices equilibrium) time path of capital accumulation Given these optimal sequences, consumption and effort can be computed from (3.1) and (3.2). It is also direct to compute variations in output, investment, real wages and real interest rates. For in (3.4). example, output variations are given by With Cobb-Douglas production, real wages are proportional to labor productivity, so that In general, optimal decisions for consumption, capital, effort, etc. depend As demonstrated in the technical appendix, on the entire sequence the time path of efficient capital accumulation may be written in the form (3.7) where 1 and 2 are complicated functions of the underlying parameters of preferences and technology. The dynamics of capital accumulation depend on the previous period’s capital stock with coefficient µ1. In addition, with time-varying total factor productivity, the optimal solution for capital accumulation depends on the current productivity level and on the entire future time path of displacements to productivity ‘discounted’ by µ2. 3.2. Transition path dynamics In order to provide a quantitative evaluation of the dynamic properties of the neoclassical model we choose a set of parameters values that match the average U.S. growth experience. The properties of the transition paths to the steady state capital stock (k) can then be numerically investigated by setting At=A for all t. In this case, the (approximately) optimal sequence of transformed capital stocks described by (3.7) reduces to the first-order with Given an initial condition difference equation k0=K0/X0, the transformed economy’s capital stock approaches its steady state value more quickly the closer µ1 is to zero. In addition, the variations © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 120 FOUNDATIONS R.G.King et al., Production, growth and business cycles I 207 in consumption, investment, output, work effort, the real wage and the real interest rate are determined according to linear relations: (3.8) where r is the steady state real interest rate, r=X/*–1. Except for πrk, the π coefficients should be interpreted as the elasticities of the flow variables with respect to deviations of the capital stock from its stationary value. The transition paths of these flow variables, therefore, are simply scaled versions of the capital stock’s transition path. In general, the values of µ1 and the π coefficients are complicated functions of the underlying parameters of the model, i.e., α, σ, δ K, β and X. 3.2.1. A fixed labor experiment Within the neoclassical model with fixed labor, variations in alter substitution over time. Table 1 summarizes the quantitative effects of varying on the adjustment parameter µ1 and the π coefficients.17 The values of the underlying parameters assume that the time interval is a quarter and are summarized in the table. Labor’s share α=0.58 is the average ratio of total employee compensation to GNP for the period 1948 to 1986; X is one plus the common trend rate of growth of output, consumption and investment, which is 1.6% per year in the post-war era.18 The value for β*=X/(1+r) is chosen to yield a return to capital of 6.5% per annum, which is the average real return to equity from 1948 to 1981.19 Finally, the depreciation rate is set at 10% per annum, which leads to a share of gross investment of 0.295. In the fixed labor model, some of the π coefficients are invariant to . The elasticities of output and real wages with respect to capital are simply determined by πyk=πwk=(1–α) which is 0.42 in our table 1 example. The value of πrk=ηk is also invariant to and takes the value –0.024. This means that output and real wages move directly with capital and real interest rates inversely with capital. In the case of log utility (=1), the table shows that the adjustment coefficient (µ1) is 0.966 which implies that one-half of any initial deviation from the stationary state is worked off in 20 quarters or 5 years. If the 17 Tedious algebra shows that πck=[(1–α)–(γXµ1–(1–δK ))(k/y)]/sc and πik=[(γXµ1– (1δK))(k/y)]/si. It is direct that πNk=0 and πyk=(1-α). Finally, µ1 is the smaller root of the quadratic equation µ2-[1/β *-scηK/ si+1]µ+1/β*. 18 Details of this computation and the data used are discussed in section 5.2. 19 Note that while β* is invariant with respect to under the assumption that β*=γX/(1+r), β is not since © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 208 Half-life is defined by the solution to rounded to the nearest integer. THE BASIC NEOCLASSIC MODEL a R.G.King et al., Production, growth and business cycles I Table 1 Effects of intertemporal substitution in consumption on near steady state dynamics (fixed labor model). 121 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 122 FOUNDATIONS R.G.King et al., Production, growth and business cycles I 209 capital stock is initially below its steady state value, then investment is above its steady state rate (πik=–0.176<0) and consumption is below its steady state rate (πck=0.670>0). Alternative values of change πck, πik and µ1 in intuitive ways. For example, when is large the representative agent is less willing to substitute intertemporally and thus desires very smooth consumption profiles. Hence, there is little reaction of consumption to a shortfall in capital (πck small). Consequently the adjustment to the steady state is slower (µ1 closer to 1.0) than when =1.0. When is small, there is more willingness to substitute consumption intertemporally and thus a given capital shortfall occasions a larger reduction in consumption. There is thus a more rapid adjustment of capital (µ1 further from 1.0) than with =1. 3.2.2. Varying work effort We are also interested in the pattern of efficient variation in work effort along the transition path, how the labor-leisure margin alters the speed of capital stock adjustment (µ1) and the responses of the price and quantity variables. To investigate these effects quantitatively, we reinstate labor as a choice variable and suppose that the utility function has the simple form u(c, L)=log(c)+llog(L). The parameter l is chosen so that stationary hours are 0.20.20 Our choice of this value is based on the average percentage of time devoted to market work in the U.S. during the period 1948–1986.21 The resulting value of µ1 is 0.953, implying a half-life of just under 14 quarters for deviations of the capital stock from its stationary level. This is a slightly more rapid pace of adjustment than the comparable fixed labor case with =1 in table 1, since work effort provides an additional margin along which agents can respond. The values of the elasticities are πck=0.617, πik=–0.629, πNk=–0.294, πyk=0.249, πwk=0.544 and πrk=–0.029. Transition paths of the key variables are plotted in fig. 1. Starting from an initially low capital stock, there is a sustained period in which output and consumption are low, but rising, while work effort and investment are high, but declining. Temporary variation in work effort is efficient even though steady state hours are invariant to growth. The economic mechanisms behind these transition paths are important. The initially low capital stock has three implications for the representative consumer in the transformed economy. First, non-human wealth is low 20 In our computations, we directly specify that N=0.20 in the linear expressions (3.1) and (3.2), noting that logarithmic utility implies zero cross elasticities and unitary own elasticities. This implicitly specifies the utility function parameter l. 21 This value is equal to the average work week as a fraction of total weekly hours for the period 1948 to 1986. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE BASIC NEOCLASSIC MODEL 210 123 R.G.King et al., Production, growth and business cycles I Fig. 1 relative to its stationary level. Second, the marginal product of labor (shadow real wage) is low relative to the stationary level. Third, the marginal product of capital (shadow real interest rate) is high relative to its stationary level. The first and third factors induce the representative consumer to work additional hours; the second factor exerts the opposite influence. With the particular preferences and technology under study, the former factors dominate, resulting in hours that are high—relative to the stationary level—along the transition path from a low initial capital stock. It is beyond the scope of this paper to undertake a detailed sensitivity analysis of how the µ and coefficients change with parameters of the environment. However, we have studied how the root µ1 depends on a list of parameter values by computing an elasticity of µ1 with respect to each parameter.22 The elasticities are quite small ranging from –0.11 for labor’s share () to –0.001 for the rate of technological progress (X–1).23 22 We thank Adrian Pagan for pushing us to conduct these experiments. The elasticity for steady state hours (N) is 0.003; for depreciation (K) is –0.03 for the intertemporal elasticity of substitution ( ) is 0.03, and for the elasticity of the marginal utility of leisure (LD2(L)/D(L)) is 0.003. 23 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 124 FOUNDATIONS R.G.King et al., Production, growth and business cycles I 211 Our conclusion is that the speed of adjustment is not highly sensitive to the choice of parameter values. 4. Real business cycles This section follows the pioneering work of Kydland and Prescott (1982) and Long and Plosser (1983) by incorporating uncertainty—in the form of temporary productivity shocks—into the basic neoclassical model. Although other aspects of the underlying economic environment are identical to those of the preceding section, the business cycle analysis is in marked contrast to the standard ‘growth theory’ analysis, in which time variation in technology is taken to be smooth, deterministic and permanent. 4.1. Linear business cycle models In principle, quantitative analyses of stochastic elements should follow Brock and Mirman’s (1972) seminal analysis of the basic neoclassical model under uncertainty. One would begin by postulating a specific stationary stochastic process for technology shocks, calculate the equilibrium laws of motion for state variables (the capital stock) and related optimal policy functions for controls (consumption, investment and work effort). It would then be natural to interpret observed business fluctuations in terms of the economy’s stationary distribution. The principle barrier to the execution of this strategy is computational. The equilibrium laws of motion for capital and for flows cannot be calculated exactly for models of interest, but must be approximated with methods that are computationally burdensome.24 Further-more, computational strategies for approximate suboptimal equilibria are not well developed. In our analysis we invoke certainty equivalence, employing a linear systems perspective. Our use of certainty equivalence methods in the study of real business cycles builds on prior work by Kydland and Prescott (1982), but the details of our procedures are different.25 An advantage of 24 Examples include Sargent (1980) and Greenwood, Hercowitz and Huffman (1986). Kydland and Prescott (1982) eliminate non-linearities in constraints (such as the production function) by substituting resource constraints into the utility function and taking a quadratic approximation to the resulting return function. We derive efficiency conditions under certainty and approximate these to obtain linear decision rules. These two procedures are equivalent for the class of models we consider when standard Taylor series approximations are used with each procedure. The only substantive difference between our approximation method and Kydland and Prescott’s is that while they search for an approximation based on a likely range of variation of the different variables, we center our linearizations on the steady state. According to Kydland and Prescott (1982, p. 1357, 11) this difference in approximation techniques has little impact on their results. Our procedure yields formulas that have a transparent economic interpretation and allows us to replicate exactly the Long and Plosser (1983) closed form solution. 25 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE BASIC NEOCLASSIC MODEL 212 125 R.G.King et al., Production, growth and business cycles I our method is that it is readily extended to the study of suboptimal dynamic equilibria, as we show in our second essay. Nevertheless, a detailed analysis of the overall accuracy of these approximation methods in a business cycle context remains to be undertaken. For the basic neoclassical model, our strategy works as follows. We develop approximate solutions for capital and other variables near the stationary point of the transformed economy as in the previous section. Then, working from a certainty equivalence perspective, we posit a particular stochastic process for and replace the sequence with its conditional expectation given information available at t. In particular, suppose that follows a first-order autoregressive process with parameter ρ. Then, given (3.7), the state dynamics are given by the linear system (4.1) where and is the state vector. Additional linear equations specify how consumption, work effort, investment, shadow prices and output depend on the state variables st. Let the vector be a vector of controls and other flow variables of interest. Then the linear equations relating flows to states are (4.2) where the coefficients are determined, as in section 3, by requiring that the shadow prices and elements of zt satisfy the linearized first-order conditions.26 26 This state space formulation (4.1) and (4.2) can be solved to obtain the vector autoregressivemoving average (ARMA) representation of the endogenous variables z. In the basic neoclassical model with persistent technology shocks (≠0), each element of zt is ARMA (2, 1) with common autoregressive but different moving average polynomials. Following Zellner and Palm (1974) and Orcutt (1948), the evolution of states can be expressed as follows where B is the backshift operator, det(I–MB) is the determinant of the 2×2 matrix defined by I–MB, and adj(I–MB) is the adjoint of I–MB. From inspection of (4.1) it is clear that, for ≠0, the determinant of (I–MB) is a second-order polynomial (1-µ1B)(1-B). There are moving average terms of at most order 1 in adj(I–MB). Further, since zt=st, the elements of zt inherit the ARMA (2, 1) structure of the state vector. The relatively simple ARMA structure of the individual elements of z is a result of the dimensionality of the state vector. In a model with many state variables the order of the polynomial det(I–MB) could become quite large, implying more complex ARMA representations for the elements of z. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 126 FOUNDATIONS R.G.King et al., Production, growth and business cycles I 213 This formulation facilitates computation of (i) impulse response functions for the system and (ii) population moments of the joint (zt, st) process. Impulse responses. Impulse response functions provide information on the system’s average conditional response to a technology shock at date t, given the posited stochastic process. The response of the system in period t+k to a technology impulse at date t+1 is Population moments. Population moments provide additional, unconditional properties of the time series generated by the model economy. We stress that although there is a single shock to the economic model under study, the dynamic character of the model means that the unconditional time series will, in general, not be perfectly correlated. The linear character of the system implies that it is relatively straightforward to calculate population moments. For example, given the variance-covariance matrix of the states, it is easy to calculate the autocovariance of z at lag j, In our analysis below, we will be concerned with how these properties of the model change as we alter parameters of preferences and technology. 4.2. Alternative parameterizations of the basic neoclassical model. We explore four alternative parameterizations of the basic neoclassical model, obtained by varying certain aspects of preferences and technology. Though far from exhaustive, these parameterizations shed some light on important aspects of neoclassical models. Table 2 summarizes the parameter values that are employed in our four versions of the neoclassical model. Throughout, as in table 1, we use production parameter values for labor’s share as α=0.58 and the growth of exogenous technical progress as (γX–1) =0.004 per quarter. In all specifications, we take the momentary utility function to be of the additively separable form, u(c, L)=log(c)+θl(L). This specification implies zero cross-elasticities (ξlc=ξcl=0) and unitary elasticity in consumption (=–ξcc=1), while leaving the elasticity of the marginal utility of leisure with respect to leisure (ξll) as a parameter to be specified. The parameter θl in all parameterizations is adjusted so as to yield a steady state value for N equal to 0.20, the average time devoted to market work in the U.S. during the period 1948–1986. In all of these formulations, the values of , γX and β combine to yield a steady state real interest rate of 6.5% per annum. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 214 Table 2 Alternative model parameterizations. THE BASIC NEOCLASSIC MODEL R.G.King et al., Production, growth and business cycles I 127 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 128 FOUNDATIONS R.G.King et al., Production, growth and business cycles I 215 Our point of departure is the parameterization of Long and Plosser (1983). The key features of this specification are additively separable, logarithmic preferences, a Cobb-Douglas production function and 100% depreciation. This specification is instructive because there is an exact closed-form solution that enables us to establish a benchmark for judging our approximation methods. The second specification alters the Long-Plosser formulation by assuming less than 100% depreciation. This alteration is sufficient to obtain stochastic properties for key variables that are more compatible with common views of economic fluctuations. We refer to this case as the ‘baseline’ model— it is closely related to the divisible labor economy studied by Hansen (1985).27 The next two experiments consider some perturbations of the elasticity of labor supply. The third parameterization uses an ‘upper bound’ labor supply elasticity from the panel data studies reviewed by Pencavel (1986). This elasticity is ten times smaller than that imposed by the logarithmic preferences of the baseline mode.28 The fourth parameterization illustrates the consequences of infinite intertemporal substitutability of leisure or, equivalently, the indivisibility of individual labor supply decisions stressed by Rogerson (1988) and Hansen (1985). 4.3. Quantitative linear business cycle models The reference point for our discussion is table 3, which summarizes the linear systems representation given in eqs. (4.1) and (4.2). That is, table 3 provides the coefficients, µ1, , πkA of the matrix M and the coefficients of the Π matrix under two assumptions about persistence of technology shocks (=0 and =0.9). Long-Plosser with complete depreciation. Applying the exact solutions found in Long and Plosser (1983), the capital stock for this parameterization evolves according to the stochastic difference equation, 27 There are at least three differences between our methodology and that employed by Hansen (1985) which make our results not directly comparable. First, we use a different linearization technique, as discussed above. Second, we compute the population movements rather than estimate them through Monte Carlo simulation. Third, we do not filter the series with the Hodrick and Prescott (1980) filter. See footnote 31 for a discussion of differences in parameter values and of the effects of the Hodrick and Prescott filter. 28 For preferences separable in consumption and leisure, the elasticity of labor supply is (1–1/N)/ξ ll, where N is the steady state fraction of time devoted to work. Thus if the elasticity of labor supply is 0.4 and N=0.20, then ξ ll=-10.0. We are reluctant to adopt this economy as our benchmark given the difficulty in interpreting the disparity between the elasticity of labor supply of women and men in the context of our representative agent economy. Furthermore, Rogerson (1988) has demonstrated that, in the presence of indivisibility in individual labor supply decisions, an economy with finite elasticity of labor supply may behave as if this elasticity were infinite. Hence, our fourth parameterization has preferences consistent with an infinite elasticity of labor supply (ξll=0). © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 216 Table 3 Parameter values of the linear system (4.1)–(4.2). THE BASIC NEOCLASSIC MODEL R.G.King et al., Production, growth and business cycles I 129 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 130 FOUNDATIONS R.G.King et al., Production, growth and business cycles I 217 (4.3) which indicates that in our approximation it should be the case that µ1= (1–) and π kA =1.0. As emphasized by Long and Plosser, (4.3) illustrates that even without long-lived commodities, capitalistic production enables agents to propagate purely transitory productivity shocks forward in time in keeping with their preferences for smooth consumption. The solutions of Long and Plosser also imply that there are simple loglinear relations for the flow variables (4.4) (4.5) In percent deviations from steady state, output, consumption, and investment all share the stochastic structure of the capital stock. Work effort, on the other hand, is constant (i.e. ,). With work effort constant, real wages (proportional to output per man hour) move just like output. With =1, interest rates are equal to the expected change in consumption Thus, in terms of (4.2), yk= ck= ik= wk=(1–), yA = cA = iA = Nk =1, and Nk = NA =0 . Finally, rk =- (1– ) and Turning to the approximate solutions reported in table 3, we see that these match the exact solutions (4.3)–(4.5) for the parameter values in table 2. For example, with =0.58, the coefficient µ1=(1–)=0.42 as required by eq. (4.1) above. Further, we see that there are two special features of this parameterization previously noted by Long and Plosser in their consideration of multi-sector, log-linear business cycle models. First, the solution involves no influence of expected future technological conditions on the properties of the endogenous variables. This conclusion follows from the observation that the linear systems coefficients linking quantity variables to technology (πkA, πcA, πNA, etc.) are invariant to the persistence () in the technology shock process. Second, the relation between work effort and the state variables (πNk and πNA) indicates that the approximation method preserves the other special implications of complete depreciation, namely that effort is unresponsive to the state of the economy (πNA=πNk=0). Fundamentally, each of these invariance results reflects a special balancing of income and substitution effects. For example, more favorable exert two offsetting effects on technology conditions accumulation: (i) an income effect (since there will be more outputs at © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE BASIC NEOCLASSIC MODEL 218 131 R.G.King et al., Production, growth and business cycles I given levels of capital input) that operates to lower saving and capital accumulation and (ii) a substitution effect (arising from an increased marginal reward to accumulation) that operates to raise saving. With complete depreciation and logarithmic utility, income and substitution effects exactly offset. With respect to real interest rates, the complete depreciation model alters the model’s also helps indicate how serial correlation in implications. The coefficient rA=(–a), so that with =0 diminishing returns predominates and an impulse to A lowers the rate of return. But with high persistence (>), interest rates rise due to the shift up in the future marginal reward to investment. Long-Plosser with realistic depreciation. Adopting a more realistic depreciation rate (K=0.025 or 10% per year) dramatically alters the properties of the basic neoclassical model. The adjustment parameter µ1 rises from 0.42 to 0.953, indicating that the capital stock adjusts more slowly. Second, kA falls from 1.0 to 0.166 when =0 and is no longer invariant to serial correlation properties of These responses can be explained in terms of the basic economics of lowering the depreciation rate. First, when there is a lower depreciation rate, it follows that there is a higher steady state capital stock and a lower output-capital ratio. As K goes from 1.0 to 0.025, y/k falls from 2.4 to 0.10. This suggests a substantial decline in the elasticity kA. Second, the change in µ1 and the sensitivity of kA to reflect implications that K has for the relative importance of wealth and intertemporal substitution effects. With lower depreciation, the intertemporal technology—linking consumption today and consumption tomorrow—becomes more linear near the stationary point.29 This means that the representative agent faces less sharply diminishing returns in intertemporal production possibilities and will choose a temporally smooth consumption profile that requires more gradual elimination of deviations of the capital stock from its stationary level (µ1 rises from 0.42 when K=1 to 0.953 when K=0.025). The depreciation rate also impinges on the relative importance of substitution and wealth effects associated with for j >0). In particular, the dominance of future shifts in technology ( the wealth effect is indicated by a comparison of purely temporary (=0) with more persistent technology shocks. Capital accumulation is less responsive to technological conditions when the shocks are more persistent (i.e., kA falls from 0.166 to 0.137 when rises from 0 to 0.9). For the same reason, more persistent technology shocks imply that consumption is more responsive ( cA=0.108 when =0 and cA=0.298 when =0.9) and 29 There is a marked decline in the elasticity of the gross marginal product of capital schedule, AD1F(k, N)+(1–K), with respect to capital. It falls from –k=0.58 to 0.023. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 132 FOUNDATIONS R.G.King et al., Production, growth and business cycles I 219 investment is less responsive ( iA=5.747 when =0 and iA=4.733 when =0.9). The persistence of shocks also has implications for the response of relative prices to technology shifts. Real wages respond more elastically, since there is a smaller variation in effort when shifts are more permanent. As in the model with complete depreciation, real interest rates respond positively to technology shifts when these are highly persistent. Altering the character of intertemporal tradeoffs also has implications for labor supply via intertemporal substitution channels. When technology shifts are purely temporary (=0), a one percent change in total factor productivity calls forth a 1.33 percent change in hours. This impact is attenuated, but not eliminated, when shifts in technology are more persistent ( NA=1.05 when =0.9). The nature of these intertemporal substitution responses is perhaps best illustrated by examining impulse response functions, which are derived from the coefficients presented in table 3. The two parts of fig. 2 contain impulse responses under our alternative assumptions about the persistence of shocks. In panel A, when technology shifts are purely temporary, intertemporal substitution in leisure is very evident. In the initial period, with positive one percent technology shock, there is a major expansion of work effort. The initial period output response is more than one-for-one with because of the expansion in work effort. The bulk of the output increase goes into investment with a smaller percentage change in consumption. In subsequent periods, after the direct effect of the technology shift has dissipated, the only heritage is a capital stock higher than its steady state value. The change in the capital stock induced by the initial period technology shock is ‘worked off’ via a combination of increased consumption and reduced effort. The impacts on output are smaller, in percentage terms, than the impacts on consumption or capital, because the transition path back toward the stationary point is associated with negative net investment and negative response of effort. This means that the response function after one period in fig. 2, panel A, is determined by the internal transition dynamics given in fig. 1. The only difference is that in fig. 2 the experiment is a positive increment to the capital stock of 0.166 instead of the negative increment of –1.0. in fig. 1. In panel B of fig. 2, when technology shifts are more persistent, the and endogenous impulse responses involve a combination of exogenous There is now a protracted period in which technology shocks dynamics serve to introduce positive comovements of hours, output, consumption and investment. The magnitudes of these responses are altered by the fact that agents understand the persistent character of technological shifts. In comparison with the case where technology shifts are purely temporary, consumption is more responsive to while effort is less. Other labor supply elasticities. First, when we restrict preferences to be consistent with an ‘upper bound’ labor supply elasticity of 0.4 for prime age males © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE BASIC NEOCLASSIC MODEL 220 R.G.King et al., Production, growth and business cycles I Fig. 2 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 133 134 FOUNDATIONS R.G.King et al., Production, growth and business cycles I 221 reported by Pencavel (1986), we obtain an economy whose dynamics are broadly similar to those of the baseline model except for the amplitude of response to technology shocks. In the case of purely temporary production shocks (=0), the elasticity of response of labor to shocks in technology ( NA) is 0.317, roughly one fourth of the value of NA for the baseline model. This reduced variability in labor is accompanied by smaller variability in consumption and investment.30 Second, when the labor supply elasticity is infinite, we have an economy that is the mirror image of the previous one in terms of amplitude of response to shocks. In the case of purely temporary shocks, the values of cA and iA are roughly 1.2 times those of the baseline model, while NA is fifty percent higher. 5. Implications for time series This section develops some of the predictions that the basic neoclassical model makes about economic time series when it is driven by a single technology shock. Using the model’s organizing principles, we also present some summary statistics for post-war quarterly U.S. time series. 5.1. Variability of components of output A major feature of economic fluctuations is the differential variability in the use of inputs (labor and capital) and in the components of output (consumption and investment). Table 4 presents some selected population moments for the four alternative parameterizations that summarize the models’ implications for relative variability. The specification with complete depreciation has implications that are readily traced to the simple structure of (4.3) and (4.4). First, output, consumption and investment have identical variances. Second, with complete depreciation, investment and capital differ only in terms of timing, so that capital and output are equally variable. When realistic depreciation is imposed (K=0.025), differences in the relative variability of the components of output are introduced. Further, these implications depend on the stochastic process for the technology shifts, since the moments of time series depend on the linear system coefficients reported in table 3 (which are dependent on the persistence parameter ). With purely temporary shocks, consumption is much less variable than output (about two tenths as variable). and investment is far more variable (more than three times as variable). Labor input is much more variable than consumption and about three fourths as variable as output. 30 A productivity shock induces intertemporal substitution of leisure by raising the productivity of current versus future labor and intratemporal substitution by increasing the opportunity cost of leisure in terms of consumption. Both the elasticity of intertemporal substitution of leisure and the elasticity of intratemporal substitution are smaller in this economy than in the baseline model. The reduction in the degree of substitution contributes to a reduced variability of consumption. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 222 Table 4 Selected population moments for four alternative parameterizations. THE BASIC NEOCLASSIC MODEL R.G.King et al., Production, growth and business cycles I 135 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 136 FOUNDATIONS R.G.King et al., Production, growth and business cycles I 223 When shifts in technology become more persistent (=0.9), there are important changes in implications for relative variabilities. Consumption is now six tenths as volatile as output, which accords with the permanent income perspective and with the altered linear systems coefficients discussed previously. Labor input is less than half as volatile as output, which fundamentally reflects diminished desirability of intertemporal substitution of effort with more persistent shocks.31 Alterations in the labor supply elasticity exert predictable effects on relative variability of labor input and output, while having relatively minor implications for the relative variability of the components of output. Relative to the baseline model, the reduction in labor supply elasticity to the level suggested by the panel data studies results in a decline of the variability of labor both in absolute terms and in relation to the variability of output. The relative volatility of the labor input in terms of output implied by the model is 0.27, roughly half of the standard deviation of hours relative to detrended output in the U.S. for the period 1948–1986.32 In table 5 we present some additional time series implications of our and exhibit baseline neoclassical model. One notable feature is that almost no serial correlation in the absence of serially correlated technology shocks. This is not true for consumption, wages or interest rates, however, which are smoother and correlated with lagged values of output. 31 The baseline model is structurally identical to the divisible labor economy studied by Hansen (1985). It differs, however, in values assigned to parameters. In our notation, Hansen’s economy involves =0.64, β*=0.99, X=1.00; N=0.33 and K=0.025. These alternative parameter values have implications for the moments reported in tables 4 and 5. Using a persistence parameter =0.90, the model’s relative volatility measures (standard deviations of variables relative to standard deviation of output) are as follows: consumption (0.62), investment (2.67) and hours (0.41). Basically, relative to table 4 these results reflect the decline in labor supply elasticity implied by N=1/3 rather than N=1/5. The contemporaneous correlations with output are as follows: consumption (0.81), investment (0.92) and hours (0.81). If we filter the population moments with the Hodrick-Prescott (HP) filter, then the relative variabilities and correlations are altered. For consumption these are (0.25) and (0.80), respectively, for investment they are (3.36) and (0.99) and for hours they are (0.55) and (0.98). These alterations occur because the effect of the HP filter is to give less weight to low frequencies, downplaying persistent but transient aspects of the series in question. [See the graph of the transfer function of the HP filter in Singleton (1988).] For example, the correlation of output at the yearly interval (lag 4) is 0.72 in the unfiltered Hansen parameterization and it is 0.08 in the filtered version. It is this sensitivity of results to filtering that makes us hesitant to undertake detailed comparisons with results reported by Hansen. 32 The inability of the model to generate a sufficiently high variation in labor when the elasticity of labor supply is restricted to be consistent with panel data studies has stimulated several extensions to the basic neoclassical model. Kydland (1984) demonstrates that introducing agent heterogeneity in the model can increase the relative volatility of the average number of hours worked with respect to the volatility of labor productivity. Rogerson (1988) establishes that, in the presence of indivisibility in individual labor supply, an economy with finite elasticity of labor supply behaves as if it had an infinite elasticity of labor supply. This motivates our interest in the fourth parameterization. As Hansen (1985), we find that in this economy labor is too volatile relative to output. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 224 Table 5 Population moments: Baseline model. THE BASIC NEOCLASSIC MODEL R.G.King et al., Production, growth and business cycles I 137 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 138 FOUNDATIONS R.G.King et al., Production, growth and business cycles I 225 5.2. Some empirical issues and observations Since the early part of this century, with the NBER studies of business cycles and economic growth under the leadership of Wesley Mitchell and Simon Kuznets, it has become commonplace for macroeconomic researchers to design models to replicate the principal features of the business cycles isolated by the NBER researchers. More recently, the development of statistical and computing technology has led individual researchers to define analogous sets of ‘stylized facts’ about economic fluctuations that models are then designed to emulate. Our perspective is that the development of stylized facts outside of a circumscribed class of dynamic models is difficult at best.33 First, models suggest how to organize time series. Further, it is frequently the case that stylized facts are sensitive to the methods of detrending or prefiltering. In this investigation we take the perspective that the basic neoclassical model has implications for untransformed macroeconomic data and not some arbitrary or prespecified transformation or component that is defined outside the context of the model [cf. Hicks (1965 p. 4)]. Although we do not perform formal statistical tests of model adequacy, the manner in which we proceed with data analysis is dictated by the models under study. We have considered deterministic labor augmenting technological change that grows at a constant proportionate rate as the source of sustained growth (trend). The neoclassical model then predicts that all quantity variables (with the exception of work effort) grow at the same rate X. The non-deterministic components of consumption, output and investment are then (5.1) where y, c and i are the steady state values in the transformed economy. Labor augmenting technical progress, log(Xt), can be expressed as the simple linear trend (5.2) Thus, in the language of Nelson and Plosser (1982), the implied time series are trend stationary. Moreover, they possess a common deterministic trend. Therefore, the model instructs us to consider deviations of the log levels of GNP, consumption and investment from a common linear trend as empirical counterparts to , and Work effort, on the other hand, possess no trend and, thus, is simply deviation of the log of hours from its mean. 33 See also Koopmans (1947) and Singleton (1988). © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE BASIC NEOCLASSIC MODEL 226 139 R.G.King et al., Production, growth and business cycles I In order to provide some perspective on the models’ properties, we summarize some of the corresponding sample moments of the U.S. time series. The series we consider are the quarterly per capita values of real GNP, consumption of nondurables and services (CNS), gross fixed investment (GFI) and average weekly hours per capita.34 Following the structure (5.1) and (5.2), we detrend the log levels of each of the first three series by computing deviations from a common estimated linear time trend. The estimated common trend, which corresponds is 0.4% per quarter.35 The real wage is the to an estimate of gross average hourly earnings of production or non-supervisory workers on non-agricultural payrolls. We chose not to study interest rates because of the well-known difficulties of obtaining measures of expected real interest rates. and are presented in fig. Plots of our empirical counterparts to 3. Their properties are summarized in table 6 in a manner analogous to the summary of the baseline model in table 5. Our sample period is the first quarter of 1948 (1948.1) to the fourth quarter of 1986 (1986.4). Deviations of output from the common deterministic trend, which are plotted as a benchmark in each of the panels in fig. 3, have a standard deviation of 5.6% and range in value from –13.0% to 10%. The sample autocorrelations in table 6 indicate substantial persistence, suggesting that there may be a non-stationary component to the series not eliminated by removing a common deterministic trend. The panels A and B show empirical counterparts to and plotted against the reference variable . Consumption and investment are highly correlated with output. Table 6 reports estimated correlation coefficients of 0.85 for consumption and 0.60 for investment over the 1948.1–1986.4 sample period. Consumption is less volatile than output, with a sample standard deviation of 3.9% (versus 5.6% for output) and a sample range of –7.8% to 7.4%. Investment is more volatile than output, with a sample standard deviation of 7.6% and sample range of –20.7% to 16.3%. Further, the autocorrelation statistics in table 6 indicate substantial serial correlation in both consumption and investment. Panel C of fig. 3 contains a plot of the empirical counterpart of per capita hours as well as that of output. This labor input measure has a standard deviation of 3.0%, with a maximum value of 6.5% and a minimum value of 34 All series are taken from the CITIBASE database. GNP, CNS and GFI are quarterly values. Population (P) is the total civilian non-institutional population 16 years of age and older. Employment (E) is total workers employed as taken from the Household Survey, Bureau of Labor Statistics. Average weekly hours of all workers (H) is also from the Household Survey. Average hours per capita is then calculated as E·H/P and averaged for the quarter. The wage rate is gross average hourly earnings of production workers. 35 This is the source of the estimate of X we use to parameterize the basic model in section 3. We choose not to impose the common trend assumption on wage rates because it involves a specific assumption about market structure. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 140 FOUNDATIONS R.G.King et al., Production, growth and business cycles I Fig. 3 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 227 THE BASIC NEOCLASSIC MODEL 228 R.G.King et al., Production, growth and business cycles I Fig. 3 (continued) © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 141 142 b All variables are taken from the National Income Accounts. Relative standard deviation of z is 229 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors FOUNDATIONS a R.G. King et al., Production, growth and business cycles I Table 6 Sample moments: Quarterly U.S. data, 1948.1–1986.4.a THE BASIC NEOCLASSIC MODEL 230 143 R.G.King et al., Production, growth and business cycles I –6.2% over the post-war period. The correlation between output and hours reported in table 6 is essentially zero! Inspection of the plot, however, appears to suggest that this relation is closer if one visually corrects for the periods in which output is on average high or low. In fact, if one splits the sample into subperiods of approximately 5 years each, the correlation between output and hours is never less than 0.30 and averages 0.77. Thus, when we permit the sample mean to vary (which is what looking at subperiods effectively does), the correlation between hours and output appears much higher.36 It is important to stress that there is no theoretical justification for looking at data in subperiods. The basic neoclassical model that we have been discussing has a single source of low frequency variation (the deterministic trend in labor productivity) which has been removed from the time series under study. The sensitivity of these results to the sample period suggests the possibility of a low frequency component not removed by the deterministic trend. This is consistent with the highly persistent autocorrelation structure of output noted above. The practice of removing low frequency variation in economic data plays an important role in empirical research on business fluctuations. NBER business cycle research has generally followed Mitchell’s division of time series into cyclical episodes, removing separate cycle averages for individual series. Our belief is that this methodology is likely to remove important low frequency aspects of the relations between time series, in a manner broadly similar to the computation of correlations over subperiods. Most modern empirical analyses of cyclical interactions have also followed the practice of removing low frequency components from actual and model-generated time series.37 Studying the impact of such low frequency filtering on economic time series generated by our baseline model, King and Rebelo (1987) find that there are major distortions in the picture of economic mechanisms presented by low frequency filtering. Among these are two that are particularly relevant to the labor-output relation. First, in the theoretical economy analyzed by King and Rebelo, application of a low frequency filter raises the correlation between output and labor input. Second, a low frequency filter dramatically reduces the correlation between output and capital. Panel D of fig. 3 contains a plot of our empirical measure of While the correlation with output is positive (0.76), it is not as strong as predicted 36 The subperiod results for the other variables are qualitatively similar to the overall sample. We have also explored the use of another hours series to insure that this finding was not an artifact of our data. Using an adjusted hours series developed by Hansen (1985), which covers only the 1955.3 to 1984.1 period, the correlation is 0.28 compared to 0.48 for our series for the same period. Breaking this shorter sample into subperiods also yields higher correlations than those for the overall period for the Hansen data. 37 For example, Kydland and Prescott (1982) filter both the data and the output of their model using a filter proposed by Hodrick and Prescott (1980). Hansen (1985) follows this practice as well. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 144 FOUNDATIONS R.G.King et al., Production, growth and business cycles I 231 by the model. Moreover, the positive correlation seems to arise primarily from the association at lower frequencies. There are two main conclusions we draw from this cursory view of the data. The first, and most important, is that the one sector neoclassical model that we use as our baseline specification is not capable of generating the degree of persistence we see in the data without introducing substantial serial correlation into the technology shocks. The second is that the data suggest the related possibility of a low frequency component not captured by the deterministic trend. This motivates our interest in models with stochastic growth in the companion essay. 6. Conclusions This paper has summarized the growth and business cycle implications of the basic neoclassical model. When driven by exogenous technical change at constant rates, the model possesses a steady state growth path under some restrictions on preferences for consumption and leisure. Although these restrictions imply that labor effort is constant in the steady state, they do not imply that effort is constant along transition paths of capital accumulation or in response to temporary technology shocks. Rather, the intertemporal substitution made feasible by capital accumulation applies to both consumption and effort in general equilibrium. When driven by highly persistent technology shocks, the basic neoclassical model is capable of replicating some stylized facts of economic fluctuations. First, the model generates procyclical employment, consumption and investment. Second, the model generates the observed rankings of relative volatility in investment, output and consumption. But along other dimensions, the basic model seems less satisfactory. In particular, the principle serial correlation in output—one notable feature of economic fluctuations— derives mainly from the persistence of technology shocks. On another level, as McCallum (1987) notes, the model abstains from discussing implications of government and the heterogeneity of economic agents. Perhaps the most intriguing possibility raised by the basic model is that economic fluctuations are just a manifestation of the process of stochastic growth. In the companion essay, we discuss current research into this possibility, along with issues concerning the introduction of government and heterogeneity. References Barro, R. and R.King, 1984, Time separable preferences and intertemporal substitution models of business cycles, Quarterly Journal of Economics 99, 817–839. Becker, R., J.Boyd III and C.Foias, 1986, The existence of Ramsey equilibrium, Working paper (University of Rochester, Rochester, NY). Boyd III, J., 1986, Recursive utility and the Ramsey problem, Working paper no. 60 (Center for Economic Research, University of Rochester, Rochester, NY). © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE BASIC NEOCLASSIC MODEL 232 145 R.G.King et al., Production, growth and business cycles I Brock, W. and L.Mirman, 1972, Optimal economic growth and uncertainty: The discounted case, Journal of Economic Theory 4, 479–513. Cass, D., 1965, Optimum growth in an aggregative model of capital accumulation, Review of Economic Studies 32, 223–240. Debreu, G., 1954, Valuation equilibrium and Pareto optimum. Proceedings of the National Academy of Sciences of the U.S.A. 38, 886–893. Greenwood, J., Z.Hercowitz and G.Huffman, 1986, Investment, capacity utilization and the real business cycle, Working paper (University of Western Ontario, London, Ont.). Hansen, G., 1985, Indivisible labor and the business cycle, Journal of Monetary Economics 16, 309–327. Hicks, J., 1965, Capital and growth (Oxford University Press, New York). Hodrick, R. and E.Prescott, 1980, Post-war U.S. business cycles: An empirical investigation. Working paper (Carnegie-Mellon University, Pittsburgh, PA). King, R., 1987, Business cycles and economic growth, Lecture notes on macroeconomics (University of Rochester, Rochester, NY). King, R. and S.Rebelo, 1987, Low frequency filtering and real business cycles, in progress (University of Rochester, Rochester, NY). Koopmans, T., 1947, Measurement without theory, Review of Economics and Statistics 29, 161–172. Koopmans, T., 1965, On the concept of optimal economic growth, in: The econometric approach to development planning (Rand-McNally, Chicago, IL). Kydland, F., 1984, Labor force heterogeneity and the business cycle, Carnegie-Rochester Conference Series on Public Policy 21, 173–208. Kydland, F. and E.Prescott, 1982, Time to build and aggregate fluctuations, Econometrica 50, 1345–1370. Long, J. and C.Plosser, 1983, Real business cycles, Journal of Political Economy 91, 1345– 1370. McCallum, B.T., 1987, Real business cycles, Unpublished manuscript (Carnegie-Mellon University, Pittsburgh, PA). Nelson, C. and C.Plosser, 1982, Trends and random walks in macroeconomic time series: Some evidence and implications, Journal of Monetary Economics 10, 139–167. Orcutt, G., 1948, A study of the autoregressive nature of the time series used for Tinbergen’s model of the economic system of the United States, 1919–1932, Journal of the Royal Statistical Society B 10, 1–45. Pencavel, J. 1986, Labor supply of men: A survey, in: Orley Ashenfelter and Richard Layard, eds., Handbook of labor economics (North-Holland, Amsterdam). Phelps, E., 1966, Golden rules of economic growth (Norton, New York). Prescott, E., 1986, Theory ahead of business cycles measurement, Carnegie-Rochester Conference Series on Public Policy 25, 11–66. Prescott, E. and R.Lucas, 1972, A note on price systems in infinite dimensional space, International Economic Review 13, 416–422. Rogerson, R., 1988, Indivisible labor, lotteries and equilibrium, Journal of Monetary Economics 21, 3–16. Romer, P. and T.Shinotsuka, 1987, The Kuhn-Tucker theorem implies the transversality condition at infinity, Unpublished paper (University of Rochester, Rochester, NY). Sargent, T., 1980, Tobin’s ‘q’ and the rate of investment in general equilibrium, CarnegieRochester Conference Series on Public Policy 12, 107–154. Singleton, K., 1988, Econometric issues in the analysis of equilibrium business cycle models, Journal of Monetary Economics, this issue. Solow, R., 1956, A contribution to the theory of economic growth, Quarterly Journal of Economics 70, 65–94. Swan, T., 1963, On golden ages and production functions, in: Kenneth Berril, ed., Economic development with special references to southeast Asia (Macmillan, London). Weitzman, M., 1973, Duality theory for infinite time horizon convex models, Management Science 19, 738–789. Zellner, A. and F.Palm, 1974, Time series analysis and simultaneous equations models, Journal of Econometrics 2, 17–54. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors Part III Some extensions © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors CHAPTER 8 149 Journal of Monetary Economics 16 (1985) 309–327. North-Holland INDIVISIBLE LABOR AND THE BUSINESS CYCLE Gary D.HANSEN* University of California, Santa Barbara, CA 93106, USA A growth model with shocks to technology is studied. Labor is indivisible, so all variability in hours worked is due to fluctuations in the number employed. We find that, unlike previous equilibrium models of the business cycle, this economy displays large fluctuations in hours worked and relatively small fluctuations in productivity. This finding is independent of individuals’ willingness to substitute leisure across time. This and other findings are the result of studying and comparing summary statistics describing this economy, an economy with divisible labor, and post-war U.S. time series. 1. Introduction Equilibrium theories of the business cycle, such as Kydland and Prescott (1982) or Lucas (1977), have been criticized for failing to account for some important labor market phenomena. These include the existence of unemployed workers, fluctuations in the rate of unemployment, and the observation that fluctuations in hours worked are large relative to productivity fluctuations. Equilibrium models have also been criticized for depending too heavily on the willingness of individuals to substitute leisure across time in response to wage or interest rate changes when accounting for the last observation. This criticism is based at least partially on the fact that micro studies using panel data on hours worked by individuals have not detected the intertemporal substitution necessary to explain the large aggregate fluctuations in hours worked [see Ashenfelter (1984)]. In this paper, a simple one-sector stochastic growth model with shocks to technology is constructed in which there is high variability in the number employed and total hours worked even though individuals are relatively unwilling to substitute leisure across time. The model differs from similar models, such as Kydland and Prescott (1982), in that a non-convexity * This paper is part of my doctoral dissertation written while a student at the University of Minnesota. I have benefited from conversations with many people including Robert King, Thomas Sargent, Christopher Sims, Neil Wallace, Sumru Altug, Patrick Kehoe, Ramon Marimon, Ian Bain, and Rody Manuelli. I owe my greatest debt, however, to my advisor, Edward Prescott. I wish to also acknowledge the Federal Reserve Bank of Minneapolis which has provided support for this research. All errors, of course, are mine. 0304–3923/85/$3.30 © 1985, Elsevier Science Publishers B.V. (North-Holland) © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 150 310 SOME EXTENSIONS G.D.Hansen, Indivisible labor and the business cycle (indivisible labor) is introduced. Indivisible labor is modeled by assuming that individuals can either work some given positive number of hours or not at all—they are unable to work an intermediate number of hours. This assumption is motivated by the observation that most people either work full time or not at all. Therefore, in my model, fluctuations in aggregate hours are the result of individuals entering and leaving employment rather than continuously employed individuals adjusting the number of hours worked, as in previous equilibrium models. This is consistent with an important feature of U.S. post-war data: most fluctuation in aggregate hours worked is due to fluctuation in the number employed as opposed to fluctuation in hours per employed worker. This is a fact that previous equilibrium theories have not tried to account for.1 Existing equilibrium models have also failed to account for large fluctuations in hours worked along with relatively small fluctuations in productivity (or the real wage). Prescott (1983), for example, finds that for quarterly U.S. time series, hours worked fluctuates about twice as much (in percentage terms) as productivity. In this paper it is shown that an economy with indivisible labor exhibits very large fluctuations in hours worked relative to productivity. This stands in marked contrast to an otherwise identical economy that lacks this non-convexity. In this economy hours worked fluctuates about the same amount as productivity.2 Equilibrium theories of the business cycle have typically depended heavily on intertemporal substitution of leisure to account for aggregate fluctuations in hours worked.3 The willingness of individuals to substitute intertemporally is measured by the elasticity of substitution between leisure in different time periods implied by an individual’s utility function. However, the theory developed here is able to account for large aggregate fluctuations in hours worked relative to productivity without requiring that this elasticity be large. This follows because the utility function of the ‘representative agent’ in our model implies an elasticity of substitution between leisure in different periods that is infinite.4 This result does not depend on the elasticity of substitution implied by the preferences of the individuals who populate the economy. Thus, the theory presented here is in principle consistent with the low estimates of this elasticity found from studying panel data [see Altonji (1984) or MaCurdy (1981)]. 1 The fact that existing equilibrium models are inconsistent with this observation has been stressed by Heckman (1983) and Coleman (1984). 2 Kydland and Prescott (1982) attempt to explain the above fact by including past leisure as an argument in the individual’s utility function so as to enhance the intertemporal substitution response to a productivity shock. However, even after introducing this feature, Kydland and Prescott were still unable to account for this observation. 3 This is true for the technology shock theories, such as Kydland and Prescott’s (1982), as well as the monetary shock theories of Lucas and Barro [see Lucas (1977)]. 4 In this model there is a crucial distinction between the utility function of the ‘representative agent’ and the utility function of an individual or household. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors INDIVISIBLE LABOR AND THE BUSINESS CYCLE G.D.Hansen, Indivisible labor and the business cycle 151 311 The paper is divided as follows: The next section provides a more detailed explanation and motivation of the indivisible labor assumption. In section 3 the artificial economies to be studied are constructed. The first is a standard stochastic growth model where labor is divisible, and the second introduces indivisible labor to that economy. The second economy is a stochastic growth version of a static general equilibrium model developed by Rogerson (1984). Lotteries are added to the consumption set (following Rogerson) which makes it possible to study a competitive equilibrium by solving a representative agent problem, as in Lucas and Prescott (1971). The addition of the lotteries also implies that the firm is providing full unemployment insurance to the workers. The fourth section explains how the equilibrium decision rules and laws of motion are calculated, as well as how the parameter values used when simulating the model were chosen. Since the representative agent’s problem is not one for which a closed form solution is available, in order to calculate decision rules a quadratic approximation of this problem is derived using the method described in Kydland and Prescott (1982). These equilibrium decision rules are a set of stochastic difference equations from which the statistical properties of the time series generated by the artificial economies can be determined. The statistics studied are a set of standard deviations and correlations discussed in section 5. In this section, the statistics computed using the artificial time series are compared to the same statistics computed using U.S. time series. Some concluding remarks are contained in section 6. 2. Motivation Existing equilibrium theories of the business cycle analyze individuals who are free to adjust continuously the number of hours worked (the ‘intensive margin’) and who are always employed. There are no individuals entering or leaving employment (the ‘extensive margin’). However, the extensive margin seems important for explaining some aspects of labor supply at both the micro and macro levels. Heckman and MaCurdy (1980), for example, discuss the importance of the extensive margin for explaining female labor supply. At the aggregate level, over half of the variation in total hours worked is due to variation in the number of individuals employed rather than variation in average hours worked by those employed. Consider the following decomposition of variance involving quarterly data: var(log Ht)=var(log ht)+var(log Nt)+2cov(log ht, log Nt), where Ht is total hours worked, ht is average hours worked, and Nt is the number of individuals at work, where all variables are deviations from trend.5 5 The data used for this analysis is available from the Bureau of Labor Statistics’ Labstat data tape. The series I used were collected from households using the Current Population Survey. For a description of the detrending method, see footnote 18. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 152 312 SOME EXTENSIONS G.D.Hansen, Indivisible labor and the business cycle Using this decomposition, 55% of the variance of Ht is due to variation in Nt, while only 20% of this variance can be directly attributed to ht. The remainder is due to the covariance term.6 Most people either work full time or not at all. This might be ascribed to the presence of non-convexities either in individual preferences for leisure or in the technology. For example, the technology may be such that the marginal productivity of an individual’s work effort is increasing during the first part of the workday or workweek, and then decreasing later on. That is, the individual faces a production function which is convex at first and then becomes concave. This could be due to individuals requiring a certain amount of ‘warm up’ time before becoming fully productive. Such a technology could induce individuals to work a lot or not at all. Another possibility is that the non-convexity is a property of individuals’ preferences. If the utility function exhibited decreasing marginal utility of leisure at low levels of leisure and increasing marginal utility at higher levels, individuals would tend to choose a low level of leisure (work a lot) or use their entire time endowment as leisure (not work at all). These preferences may be interpreted as ‘indirect’ preferences which reflect costs associated with working each period, such as driving a long distance to work or enduring the hassle of putting on a suit and tie. Bearing these fixed costs makes an individual less likely to choose to work only half a day. In this paper the non-convexity is assumed to be a property of preferences.7 However, to make the model tractable, the non-convexity introduced—indivisible labor—is an extreme version of the non-convexity described above. Individuals are assumed to have preferences that are defined only at two levels of leisure— one level corresponding to working full time and the other corresponding to not working at all. This is modeled by assuming that the consumption possibilities set consists of only two levels of leisure. This assumption implies that an individual can only adjust along the extensive margin. Of course fluctuations along both the extensive and intensive margins are observed in the actual economy, as the above evidence indicates. However, by studying two economies—one that exhibits fluctuations only along the intensive margin and another with fluctuations only along the extensive margin—we can determine the importance of non-convexities for explaining labor variability in business cycles. If it turns out that both economies exhibit the same cyclical behavior, then it seems likely that a model that incorporated both margins would also exhibit similar behavior. In fact, non-convexities of this 6 Coleman (1984) comes to a similar conclusion using establishment data. One advantage of modeling the non-convexity as a feature of the technology is that it would likely explain why part-time workers are paid less than full-time workers, in addition to accounting for features of the data discussed in this paper. 7 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors INDIVISIBLE LABOR AND THE BUSINESS CYCLE G.D.Hansen, Indivisible labor and the business cycle 153 313 sort could probably be safely abstracted from when studying business cycle phenomena. However, it happens that the two models have very different implications and that the non-convexity improves our ability to account for U.S. aggregate time series data. 3. Two economies 3.1. A one-sector stochastic growth model with divisible labor The economy to be studied is populated by a continuum of identical infinitely lived households with names on the closed interval [0, 1]. There is a single firm with access to a technology described by a standard Cobb-Douglas production function of the form (1) where labor (ht) and accumulated capital (kt) are the inputs and t is a random shock which follows a stochastic process to be described below. Agents are assumed to observe t before making any period t decisions. The assumption of one firm is made for convenience. Since the technology displays constant returns to scale—implying that firms make zero profit in equilibrium—the economy would behave the same if there were many firms. Output, which is produced by the firm and sold to the households, can either be consumed (ct) or invested (it), so the following constraint must be satisfied: (2) The law of motion for the capital stock is given by (3) where δ is the rate of capital depreciation. The stock of capital is owned by the households who sell capital services to the firm. The technology shock is assumed to follow a first-order Markov process. In particular, t obeys the following law of motion: (4) where the εt’s are iid with distribution function F. This distribution is assumed to have a positive support with a finite upper bound, which guarantees that output will always be positive. By requiring F to have mean 1-γ, the unconditional mean of t is equal to 1. This technology shock is motivated by the fact that in post-war U.S. time series there are changes in output (GNP) that can not be accounted for by © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 154 314 SOME EXTENSIONS G.D.Hansen, Indivisible labor and the business cycle changes in the inputs (capital and labor). We follow Solow (1957) and Kydland and Prescott (1982) in interpreting this residual as reflecting shocks to technology. Households in this economy maximize the expected value of where 0< β <1 is the discount factor and ct and lt are consumption and leisure in period t, respectively. The endowment of time is normalized to be one, so lt=1-ht. Utility in period t is given by the function (5) We now have a complete specification of the preferences, technology, and stochastic structure of a simple economy where individuals are able to supply any level of employment in the interval [0, 1]. Each period three commodities are traded: the composite output commodity, labor, and capital services. It is possible to consider only this sequence of spot markets since there is no demand for intertemporal risk sharing which might exist if households were heterogeneous. Households solve the following problem, where wt is the wage rate at time t and rt is the rental rate of capital: (6) Agents are assumed to make period t decisions based on all information available at time t (which includes rt and wt). They have rational expectations in that their forecasts of future wages and rental rates are the same as those implied by the equilibrium laws of motion. The first-order conditions for the firm’s profit maximization problem imply that the wage and rental rate each period are equal to the marginal productivity of labor and capital, respectively. Since there are no externalities or other distortions present in this economy, the equal-weight Pareto optimum can be supported as a competitive equilibrium. Since agents are homogeneous, the equal-weight Pareto optimum is the solution to the problem of maximizing the expected welfare of the representative agent subject to technology constraints. This problem is the following: (7) © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors INDIVISIBLE LABOR AND THE BUSINESS CYCLE G.D.Hansen, Indivisible labor and the business cycle 155 315 The state of the economy in period t is described by kt and t. The decision variables are ht, ct, and it. This problem can be solved using dynamic programming techniques.8 This requires finding the unique continuous function (where S is the state space) that satisfies Bellman’s equation (primes denote next period values) (8) where the maximization is over c and h and is subject to the same constraints as (7). The value function, V(k, ), is the maximum obtainable expected return over all feasible plans. It turns out that since the utility function is concave and the constraint set convex, the value function is also concave. This implies that the problem (8) is a standard finite-dimensional concave programming problem. Unfortunately, this problem is not one which can be solved analytically. There is no known explicit functional form for the value function, V. In principle this problem could be solved using numerical methods [see Bertsekas (1976)], but a cheaper method—which does enable one to solve for closed form decision rules—is to approximate this problem by one which consists of a quadratic objective and linear constraints, as in Kydland and Prescott (1982). This method will be explained briefly in section 4. 3.2. An economy with indivisible labor The assumption of indivisible labor will now be added to the above stochastic growth model. This will give rise to an economy where all variation in the labor input reflects adjustment along the extensive margin. This differs from the economy described above where all variation in the labor input reflects adjustment along the intensive margin. In addition, the utility function of the ‘representative agent’ for this economy will imply an elasticity of substitution between leisure in different periods that is infinite and independent of the elasticity implied by the utility function of the individual households. Indivisibility of labor is modeled by restricting the consumption possibilities set so that individuals can either work full time, denoted by h0, or not at all.9 8 For a detailed presentation of dynamic programming methods, see Lucas, Prescott and Stokey (1984). 9 This is consistent with the interpretation given in section 2. An alternative interpretation of indivisible labor assumes that households can work one of two possible (non-zero) number of hours, h1 or h2. This interpretation is consistent with an environment where each household consists of two individuals, at least one of whom works at all times. When only one member works, the household is working h1 hours, and when both members work the household is working h2 hours. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 156 316 SOME EXTENSIONS G.D.Hansen, Indivisible labor and the business cycle In order to guarantee [using Theorem 2 of Debreu (1954)] that the solution of the representative agent’s problem can be supported as a competitive equilibrium, it is necessary that the consumption possibilities set be convex. However, if one of the commodities traded is hours worked (as in the above model), the consumption possibilities set will be non-convex. To circumvent this problem, we convexify the consumption possibilities set by requiring individuals to choose lotteries rather than hours worked, following Rogerson (1984).10 Thus, each period, instead of choosing manhours, households choose a probability of working, αt.11 A lottery then determines whether or not the household actually works. After changing the economy in this manner, we make it possible for the competitive equilibrium to be derived by solving a concave programming problem, just as for the economy with divisible labor. The new commodity being introduced is a contract between the firm and a household that commits the household to work h0 hours with probability αt. The contract itself is being traded, so the household gets paid whether it works or not. Therefore, the firm is providing complete unemployment insurance to the workers. Since all households are identical, all will choose the same contract—that is, the same αt. However, although households are ex ante identical, they will differ ex post depending on the outcome of the lottery: a fraction αt of the continuum of households will work and the rest will not.12 Using (5), expected utility in period t is given by αt(log ct+A log(1-h0))+ (1αt)(log c t+A log 1). 13 This simplifies to the following function (9) 10 In Rogerson’s paper, a static economy with indivisible labor is studied and lotteries are introduced to solve the problem introduced by this non-convexity. Readers may wish to consult Rogerson’s paper for a rigorous general equilibrium formulation of this type of model. 11 Adding lotteries to the consumption set increases the choices available to households when labor is indivisible. If lotteries were not available, households would only be able to choose to not work (corresponding to α=0) or to work h0 (corresponding to α=1). Therefore, adding lotteries can only make individuals better off. 12 The lottery involves drawing a realization of a random variable zt from the uniform distribution on [0, 1]. Each individual i苸[0, 1] is now ‘renamed’ according to the following rule: The amount worked by agent x in period t is equal to This provides a mechanism for dividing the continuum of agents into two subsets, one where each individual works zero hours and another where individuals work h0. The first will have measure (1-αt) and the other measure α t. This follows from the easily verified fact that Prob[xt(i, z)ⱕ1-αt] is equal to 1-αt for each i. 13 This uses the fact that, since preferences are separable in consumption and leisure, the consumption level chosen in equilibrium is independent of whether the individual works or not. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors INDIVISIBLE LABOR AND THE BUSINESS CYCLE G.D.Hansen, Indivisible labor and the business cycle 157 317 Since a fraction αt of households will work h0 and the rest will work zero, per capita hours worked in period t is given by (10) The other features of this economy are exactly the same as for the economy with divisible labor. These include the technology and the description of the stochastic process for the technology shock. These features are described by eqs. (1) through (4). Firms in the economy, as in the previous economy, will want to employ labor up to the point where fh( t, kt, ht )=wt. However, due to the fact that lottery contracts are being traded, households are not paid for the time they actually spend working, but are instead paid for the expected amount of time spent working. This implies that each worker is paid as if he worked ht [as defined in (10)] rather than for the amount he actually does work. Therefore, the budget constraint of a typical household differs from the budget constraint for the economy where labor is divisible (6) and is given by (11) Thus, the problem solved by a typical household is (12) This problem is equivalent to the problem solved by households in a slightly different economy where agents trade man-hours and actuarially fair insurance contracts, rather than the type of contracts traded in the economy studied here. In this alternative economy, which is described in more detail in the appendix, households only get paid for the time they actually spend working. However, if a household has purchased unemployment insurance, it will receive compensation if the lottery determines that the household does not work. In the appendix it is shown that households will choose to insure themselves fully. Therefore, in equilibrium, the households will have full unemployment insurance, just like the households populating the economy described in this section. This implies that the equilibrium allocations for these two economies are the same. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 158 318 SOME EXTENSIONS G.D.Hansen, Indivisible labor and the business cycle The following is the representative agent’s problem that must be solved to derive the equilibrium decision rules and laws of motion: (13) Like problem (7), this is a standard concave discounted dynamic programming problem. The state of the economy in period t is described by kt and t. The decision variables are αt, ct, and it. A key property of this economy is that the elasticity of substitution between leisure in different periods for the ‘representative agent’ is infinite. To understand this result, first substitute ht=1-lt into (10) and solve for αt. After substituting this expression for αt into (9) one obtains the following utility function for the representative agent (ignoring the constant term): (14) where B=-A(log(l-h0))/h0. Since this utility function is linear in leisure it implies an infinite elasticity of substitution between leisure in different periods. This follows no matter how small this elasticity is for the individuals populating the economy. Therefore, the elasticity of substitution between leisure in different periods for the aggregate economy is infinite and independent of the willingness of individuals to substitute leisure across time.14 4. Solution method and calibration The problems (7) and (13) are not in the class of problems for which it is possible to solve analytically for decision rules. This special class of problems includes those with quadratic objectives and linear constraints, as well as some other structures. For this reason, approximate economies are studied for which the representative agent’s problem is linear-quadratic [see Kydland and Prescott (1982)]. It is then possible to obtain explicit decision rules for these approximate economies. By making appropriate substitutions, one can express problems (7) and (13) as dynamic optimization problems with decision variables it and ht and state variables t and kt. The constraints for these problems are linear although the 14 The fact that in this type of model the representative agent’s utility function is linear in leisure was originally shown by Rogerson (1984) for his model. This result depends, however, on the utility function being additively separable across time. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors INDIVISIBLE LABOR AND THE BUSINESS CYCLE G.D.Hansen, Indivisible labor and the business cycle 159 319 objective functions are non-linear. For each of these problems, Kydland and Prescott’s procedure is used to construct a quadratic approximation of the objective function to be accurate in a neighborhood of the steady state for the appropriate model after the technology shock has been set equal to its unconditional mean of one.15 The reader may consult Kydland and Prescott (1982) for details on the algorithm used for forming these approximations.16 To actually compute these quadratic approximations, solve for an equilibrium, and generate artificial time series, it is necessary to choose a distribution function, F, and specific parameter values for θ, δ, β, A, γ, and h0. Kydland and Prescott (1982, 1984) follow a methodology for choosing parameter values based on evidence from growth observations and micro studies. This methodology will also be followed here. In fact, since they study a similar economy, some of the above parameters (θ, δ, β) also appear in their model. This enables me to draw on their work in selecting values for these parameters, thereby making it easier to compare the results of the two studies. The parameter θ corresponds to capital’s share in production. This has been calculated using U.S. time series data by Kydland and Prescott (1982, 1984) and was found to be approximately 0.36. The rate of depreciation of capital, δ, is set equal to 0.025 which implies an annual rate of depreciation of 10 percent. Kydland and Prescott found this to be a good compromise given that different types of capital depreciate at different rates. The discount factor, β, is set equal to 0.99, which implies a steady state annual real rate of interest of four percent. The parameter A in the utility function (5) is set equal to 2. This implies that hours worked in the steady state for the model with divisible labor is close to 1/3. This more or less matches the observation that individuals spend 1/3 of 15 Let the steady states for the certainty version of these models be denoted by the variable’s symbol without any subscript. Eq. (3) implies that investment in the steady state is given by i=δk. Expressions for k and h can be determined by deriving the Euler equations for the appropriate representative agent problem and setting ht=h, kt=k, and it=i=δk for all t. For both economies, the steady state capital stock is given by Hours worked in the steady state for the economy with divisible labor is given by h=(1-θ)× (ρ+δ)/[3(ρ+δ)-θ(ρ+3δ)]; and for the economy with indivisible labor, h=(1-θ )(ρ+δ)/[ (ρ+δ–θ δ)] where =-A[log(1-h0)]/h0. 16 Kydland and Prescott’s method for approximating this problem requires choosing a vector of average deviations, which determines the size of the neighborhood around the steady state within which the approximation is accurate. The four components of z are average deviations from trend of the four variables, xt=(t, kt, it, ht), as found in U.S. time series data. This implies that along those dimensions where there is more variability, the approximation will be accurate in a larger neighborhood around the steady state . For the exercise carried out in this paper reflecting the average standard deviations of these series as reported in the next section. Although attention was paid to specifying this vector in a reasonable way, it turns out that the results are not altered when the zt components are decreased by a factor of ten. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 160 SOME EXTENSIONS 320 G.D.Hansen, Indivisible labor and the business cycle their time engaged in market activities and 2/3 of their time in non-market activities. To determine the parameter h0, I set the expressions for hours of work in the steady state for the two models equal to each other. Since steady state hours worked in the model with divisible labor is fully determined by the parameters θ, δ, A, and β for which values have already been assigned (see footnote 15), it is possible to solve for h0. This implies a value for h0 of 0.53. The distribution function F along with the parameter γ determine the properties of the technology shock, t. The distribution of εt is assumed to be log normal with mean (1- γ), which implies that the unconditional mean of t is 1. The parameter γ is set equal to 0.95 which is consistent with the statistical properties of the production function residual.17 The standard deviation of εt, σε, is difficult to measure from available data since this number is significantly affected by measurement error. A data analysis suggests that e could reasonably be expected to lie in the interval [0.007, 0.01]. A value of 0.007, for example, would imply that a little over half of the variability in εt is being attributed to measurement error, which is probably not unreasonable. The actual value used for the simulations in this paper is 0.00712. This particular value was chosen because it implies that the mean standard deviation of output for the economy with indivisible labor is equal to the standard deviation of GNP for the U.S. economy (see next section). All parameters of the two models have now been determined. We are now ready to study and compare the statistical properties of the time series generated by these two models. 5. Results For the purposes of this study, the statistical properties of the economies studied are summarized by a set of standard deviations and correlations with output that are reported in table 1. The statistics for the U.S. economy are reported in the first two columns of the table. Before these statistics were calculated, the time series were logged and deviations from trend were computed. Detrending was necessary because the models studied abstract from growth. The data were logged so that standard deviations can be interpreted as mean percentage deviations from 17 The production function residual is measured, using U.S. time series, by log t=log yt-θ log kt-(1-θ)log h t , where data on GNP, capital stock (nonresidential equipment and structures), and hours worked is obtained from a standard econometric data base. The first-order autocorrelation coefficient for t is about 0.95, indicating high serial correlation in this series. The parameter θ was assumed to be equal to 0.36 for calculating this residual. A more detailed study of the statistical properties of this technology shock is planned but has not yet been carried out. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors INDIVISIBLE LABOR AND THE BUSINESS CYCLE G.D.Hansen, Indivisible labor and the business cycle 161 321 Table 1 Standard deviations in percent (a) and correlations with output (b) for U.S. and artificial economies. a The U.S. time series used are real GNP, total consumption expenditures, and gross private domestic investment (all in 1972 dollars). The capital stock series includes nonresidential equipment and structures. The hours series includes total hours for persons at work in nonagricultural industries as derived from the Current Population Survey. Productivity is output divided by hours. All series are seasonally adjusted, logged and detrended. b The standard deviations and correlations with output are sample means of statistics computed for each of 100 simulations. Each simulation consists of 115 periods, which is the same number of periods as the U.S. sample. The numbers in parentheses are sample standard deviations of these statistics. Before computing any statistics each simulated time series was logged and detrended using the same procedure used for the U.S. time series. trend. The ‘detrending’ procedure used is the method employed by Hodrick and Prescott (1980).18 Since much of the discussion in this section centers on the variability of hours worked and productivity (output divided by hours worked), some discussion of the hours series is appropriate. The time series for hours worked used in constructing these statistics is derived from the Current Population Survey, which is a survey of households. This series was chosen in preference to the other available hours series which is derived from the establishment survey. The hours series based on the household survey is more comprehensive than 18 This method involves choosing smoothed values solve the following problem: for the series which where >0 is the penalty on variation, where variation is measured by the average squared second difference. A larger value of implies that the resulting {s t} series is smoother. Following Prescott (1983), I choose =1600. Deviations from the smooth series are formed by taking dt=xt-st. This method is used in order to filter out low frequency fluctuations. Although other methods (spectral techniques, for example) are available, this method was chosen because of its simplicity and the fact that other methods lead to basically the same results [see Prescott (1983)]. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 162 322 SOME EXTENSIONS G.D.Hansen, Indivisible labor and the business cycle the establishment series since self-employed workers and unpaid workers in family-operated enterprises are included. Another advantage is that the household series takes into account only hours actually worked rather than all hours paid for. That is, it doesn’t include items such as paid sick leave. A disadvantage is that the household series begins in the third quarter of 1955, which prevented me from using data over the entire post-war period. Sample distributions of the summary statistics describing the behavior of the artificial economies were derived using Monte Carlo methods. The model was simulated repeatedly to obtain many samples of artificially generated time series. Each sample generated had the same number of periods (115) as the U.S. time series used in the study. Before any statistics were computed, the data were logged and the same filtering procedure applied to the U.S. data was applied to these time series. One hundred simulations were performed and sample statistics were calculated for each data set generated. The sample means and standard deviations of these summary statistics are reported in the last four columns of table 1. When comparing the statistics describing the two artificial economies, one discovers that the economy with indivisible labor displays significantly larger fluctuations than the economy with divisible labor. This shows that indivisible labor increases the volatility of the stochastic growth model for a given stochastic process for the technology shock. In fact, it is necessary to increase ε by 30 percent (from 0.00712 to 0.00929) in order to increase the standard deviation of output for the divisible labor economy so that it is equal to the standard deviation of GNP for the actual economy, which is 1.76. It is still the case that 0.00929 is in the interval suggested by the data (see paragraph on measuring ε in the previous section). However, since it is likely that there is significant measurement error in our empirical estimate of the production function residual, one should prefer the lower value of ε. Another conclusion drawn from studying this table is that the fluctuations in most variables are larger for the actual economy than for the indivisible labor economy. It is my view that most of this additional fluctuation (except in the case of the consumption series) is due to measurement error. Work in progress by the author attempts to correct for measurement error in the hours series (and hence some of the measurement error in the productivity series).19 Preliminary findings seem to suggest that the above hypothesis is correct. In addition, the fact that the consumption series fluctuates much more in the actual economy than in the artificial economy can probably be explained by the fact that nothing corresponding to consumer durables is modeled in the economies studied here. 19 The work referred to is a chapter of my dissertation. Copies will soon be available upon request. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors INDIVISIBLE LABOR AND THE BUSINESS CYCLE G.D.Hansen, Indivisible labor and the business cycle 163 323 Perhaps the most significant discovery made by examining table 1 is that the amount of variability in hours worked relative to variability in productivity is very different for the two model economies. This relative variability can be measured by the ratio of the standard deviation in hours worked to the standard deviation in productivity. For the economy with indivisible labor, this ratio is 2.7, and for the economy without this feature the ratio is not significantly above 1.20 For the U.S. economy the ratio is equal to 1.4, which is between these two values. As explained in the introduction, accounting for the large variability in hours worked relative to productivity has been an open problem in equilibrium business cycle theory. Kydland and Prescott (1982) study a version of the stochastic growth model where labor is divisible and the utility function of individuals is non-time-separable with respect to leisure. This non-timeseparability property is introduced to make leisure in different periods better substitutes. However, this feature enables these authors to report a value for this ratio of only 1.17, which is still much too low to account for the fluctuations found in U.S. data. On the other hand, the economy with indivisible labor studied here has exactly the opposite problem Kyland and Prescott’s model has. The ratio implied by this model is much larger than the ratio implied by the data. However, this should not be surprising. In fact, it would be bothersome if this were not the case. After all, we do observe some adjustment along the intensive margin in the real world. Examples include workers who work overtime in some periods and not in others or salesmen who work a different number of hours each day. Since indivisible labor implies that all fluctuations are along the extensive margin, one would expect—even without looking at statistics calculated from the data—that the ratio discussed above should be somewhere between the one implied by an indivisible labor economy and a divisible labor economy. 6. Conclusion A dynamic competitive equilibrium economy with indivisible labor has been constructed with the aim of accounting for standard deviations and correlations with output found in aggregate economic time series. Individuals in this economy are forced to enter and exit the labor force in response to technology shocks rather than simply adjusting the number of hours worked while remaining continuously employed. Therefore, this is an equilibrium model which exhibits unemployment (or employment) fluctuations in response to aggregate shocks. Fluctuations in employment seem important for fluctuations This ratio is still not significantly different from one even when ε is increased to 0.00929. 20 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 164 324 SOME EXTENSIONS G.D.Hansen, Indivisible labor and the business cycle in hours worked over the business cycle since most of the variability in total hours is unambiguously due to variation in the number employed rather than hours per employed worker. An important aspect of this economy is that the elasticity of substitution between leisure in different periods for the aggregate economy is infinite and independent of the elasticity of substitution implied by the individuals’ utility function. This distinguishes this model, or any Rogerson (1984) style economy, from one without indivisible labor. These include the model presented in section 3.1 and the economy studied by Kydland and Prescott (1982). In these divisible labor models, the elasticity of substitution for the aggregate economy is the same as that for individuals. This feature enables the indivisible labor economy to exhibit large fluctuations in hours worked relative to fluctuations in productivity. Previous equilibrium models of the business cycle, which have all assumed divisible labor, have been unsuccessful in accounting for this feature of U.S. time series. This is illustrated in this paper by showing that a model with divisible labor fails to exhibit large fluctuations in hours worked relative to productivity while the model with indivisible labor displays fluctuations in hours relative to productivity which are much larger than what is observed. This seems to indicate that a model which allowed for adjustment along both the extensive margin as well as the intensive margin would have a good chance for successfully confronting the data. In conclusion, this study demonstrates that non-convexities such as indivisible labor may be important for explaining the volatility of hours relative to productivity even when individuals are relatively unwilling to substitute leisure across time. They are also useful for increasing the size of the standard deviations of all variables relative to the standard deviation of the technology shock. Therefore, a smaller size shock is sufficient for explaining business cycle fluctuations than was true for previous models such as Kydland and Prescott’s (1982). In addition, these non-convexities make it possible for an equilibrium model of the business cycle to exhibit fluctuations in employment. Therefore, non-convexities will inevitably play an important role in future equilibrium models of the cycle. Appendix: A market for unemployment insurance The purpose of this appendix is to show that the equilibrium of the economy presented in section 3.2 is equivalent to the equilibrium of an economy where labor is still indivisible but households are able to purchase any amount of unemployment insurance they choose. In the original economy, agents are assumed to buy and sell contracts which specify a probability of working in a given period as opposed to buying and selling hours of work. A lottery determines which households must work and which do not. A household is © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors INDIVISIBLE LABOR AND THE BUSINESS CYCLE G.D.Hansen, Indivisible labor and the business cycle 165 325 paid according to the probability that it works, not according to the work it actually does. In other words, the firm is automatically providing full unemployment insurance to the households. In this appendix, households choose a probability of working each period and a lottery is held to determine which households must work, just as in the original economy. Also, preferences, technology, and the stochastic structure are exactly the same as for the original model. However, this economy is different in that households only get paid for the work they actually do— unemployed individuals get paid nothing by the firm. But, the household does have access to an insurance market which preserves the complete markets aspect of the original model. It is shown below that the equilibrium of this economy is equivalent to that of the original economy since individuals will choose to be fully insured in equilibrium. This is shown by proving that the problem solved by households is the same as the problem solved by households (12) in the original model. The problem solved by the households can be described as follows: Each period, households choose a probability of working, αt, a level of unemployment compensation, yt, and consumption and investment contingent on whether the household works or not, cst and ist (s=1, 2). These are chosen to solve the following dynamic programming problem (primes denote next period values): (A.1) subject to (A.2) (A.3) (A.4) The function V(, K, k) is the value function which depends on the household’s state. The state vector includes the capital owned by the household, plus the economy wide state variables and K, where K is the per capita capital stock.21 The functions w(, K) and r(, K) are the wage rate and rental rate 21 Since we are allowing households to choose any level of unemployment insurance they wish, we have to allow for the heterogeneity that may come about because different households will have different income streams. This is why the distinction is made between the per capita capital stock, K, and the households accumulated capital stock, k. However, this heterogeneity will disappear in equilibrium since all households will choose full insurance, so K=k in equilibrium. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 166 326 SOME EXTENSIONS G.D.Hansen, Indivisible labor and the business cycle of capital respectively, and p(α) is the price of insurance, which is a function of the probability that the household works. Also, since individuals’ preferences are the same as for the original model, u(c)=log c and v(l)=A log l. The insurance company in this economy maximizes expected profits which are given by p(α)y-(1-α)y. That is, the firm collects revenue p(α)y and pays y with probability 1-α. To guarantee that profits are bounded, p(α)=(1-α). Therefore, the price the household must pay for insurance equals the probability that the household will collect on the insurance. One can now substitute this expression for p into constraints (A.2) and (A.3). After eliminating the constraints by substituting out is and cs (s=1, 2), one can write the following first-order necessary conditions for and y: (A.5) (A.6) Eq. (A.6) implies, given the strict concavity of u, that c 1=c 2 . This plus eq. (A.5) imply that This, in turn, implies that i1=i2. Therefore, the lefthand sides of eqs. (A.2) and (A.3) are identical. Since these constraints will be binding in equilibrium, y will be chosen so that the right-hand sides are equal as well. This means that y=wh0 in equilibrium. That is, households will choose to insure themselves fully. This has the implication that all households will choose the same sequence of capital stocks, so K=k. Substituting these results into the household’s optimization problem (A.1) yields the following problem: Households choose c, i, k′ , and α to (A.7) This problem is identical to problem (12). Therefore, the equilibrium allocation for the original economy, where the firm provides full unemployment insurance to workers by assumption, is equivalent to the equilibrium allocation for an economy where households get paid by the firm only for work done but have access to a risk-neutral insurance market. This result, of course, depends crucially on the probability α being publicly observable and the contract being enforceable. That is, it must be the case that the agent announces the same α to both the firm and the insurance company, and if the agent loses the lottery (that is, has to work) this is known by all parties. For example, this result © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors INDIVISIBLE LABOR AND THE BUSINESS CYCLE G.D.Hansen, Indivisible labor and the business cycle 167 327 would not hold if α depended on some underlying choice variable like effort that was not directly observed by the insurance company. In this case a difficult moral hazard problem would arise. References Altonji, J.G., 1984, Intertemporal substitution in labor supply: Evidence from micro data, Unpublished manuscript (Columbia University, New York). Ashenfelter, O., 1984, Macroeconomic analyses and microeconomic analyses of labor supply, Carnegie-Rochester Conference Series on Public Policy 21, 117–156. Bertsekas, D.P., 1976, Dynamic programming and stochastic control (Academic Press, New York). Coleman, T.S., 1984, Essays on aggregate labor market business cycle fluctuations, Unpublished manuscript (University of Chicago, Chicago, IL). Debreu, G., 1954, Valuation equilibrium and Pareto optimum, Proceedings of the National Academy of Sciences 40, 588–592. Heckman, J.J., 1984, Comments on the Ashenfelter and Kydland papers, Carnegie-Rochester Conference Series on Public Policy 21, 209–224. Heckman, J.J. and T.E.MaCurdy, 1980, A life cycle model of female labor supply, Review of Economic Studies 47, 47–74. Hodrick, R.J. and E.C.Prescott, 1980, Post-war U.S. business cycles: An empirical investigation. Working paper (Carnegie-Mellon University, Pittsburgh, PA). Kydland, F.E. and E.C.Prescott, 1982, Time to build and aggregate fluctuations, Econometrica 50, 1345–1370. Kydland, F.E. and E.C.Prescott, 1984, The workweek of capital and labor. Unpublished manuscript (Federal Reserve Bank of Minneapolis, Minneapolis, MN). Lucas, R.E., Jr., 1977, Understanding business cycles, Carnegie-Rochester Conference Series on Public Policy 5, 7–29. Lucas, R.E., Jr. and E.C.Prescott, 1971, Investment under uncertainty, Econometrica 39, 659–681. Lucas, R.E., Jr., E.C.Prescott and N.L.Stokey, 1984, Recursive methods for economic dynamics, Unpublished manuscript (University of Minnesota, Minneapolis, MN). MaCurdy, T.E., An empirical model of labor supply in a life-cycle setting, Journal of Political Economy 89, 1059–1085. Rogerson, R., Indivisible labour, lotteries and equilibrium, Unpublished manuscript (University of Rochester, Rochester, NY). Prescott, E.C., Can the cycle be reconciled with a consistent theory of expectations? or a progress report on business cycle theory, Unpublished manuscript (Federal Reserve Bank of Minneapolis, Minneapolis, MN). Solow, R.M., Technical change and the aggregate production function, The Review of Economics and Statistics 39, 312–320. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 168 CHAPTER 9 Federal Reserve Bank of Minneapolis Quarterly Review Spring 1992 The Labor Market in Real Business Cycle Theory* Gary D.Hansen Professor of Economics University of California, Los Angeles Randall Wright Senior Economist Research Department Federal Reserve Bank of Minneapolis and Professor of Economics University of Pennsylvania The basic objective of the real business cycle research program is to use the neoclassical growth model to interpret observed patterns of fluctuations in overall economic activity. If we take a simple version of the model, calibrate it to be consistent with long-run growth facts, and subject it to random technology shocks calibrated to observed Solow residuals, the model displays short-run cyclical behavior that is qualitatively and quantitatively similar to that displayed by actual economies along many important dimensions. For example, the model predicts that consumption will be less than half as volatile as output, that investment will be about three times as volatile as output, and that consumption, investment, and employment will be strongly positively correlated with output, just as in the postwar U.S. time series.1 In this sense, the real business cycle approach can be thought of as providing a benchmark for the study of aggregate fluctuations. In this paper, we analyze the implications of real business cycle theory for the labor market. In particular, we focus on two facts about U.S. time series: the fact that hours worked fluctuate considerably more than productivity and the fact that the correlation between hours worked and productivity is close to zero.2 These facts and the failure of simple real business cycle models to account for them have received considerable attention in the literature. [See, for example, the extended discussion by Christiano and Eichenbaum (1992) and the references they provide.] Here we first document the facts. We then present a baseline real business cycle model (essentially, the divisible labor The Editorial Board for this paper was V.V.Chari, Preston J.Miller, Richard Rogerson, and Kathleen S.Rolfe. model in Hansen 1985) and compare its predictions with the facts. We then consider four extensions of the baseline model that are meant to capture features of the world from which this model abstracts. Each of these extensions has been discussed in the literature. However, we analyze them in a unified framework with common functional forms, parameter values, and so on, so that they can be more easily compared and evaluated in terms of how they affect the model’s ability to explain the facts. The standard real business cycle model relies exclusively on a single technology shock to generate fluctuations, so the fact that hours worked vary more than productivity implies that the short-run labor supply elasticity must be large. The first extension of the model we consider is to recognize that utility may depend not only on leisure today but also on past leisure; this possibility leads us to introduce nonseparable preferences, as in Kydland and Prescott 1982. 3 This extension of the baseline model has the effect of increasing the relevant elasticity, by making households more willing to substitute leisure in one period for leisure in another period in response to short-run productivity changes. At the same time, with these * This paper is also available in Spanish in Cuadernos Economicos de ICE, a quarterly publication of the Ministerio de Economía y Hacienda. The paper appears here with the permission of that publication’s editor, Manuel Santos. 1 These properties are also observed in other countries and time periods. See Kydland and Prescott 1990 for an extended discussion of the postwar U.S. data, and see Blackburn and Ravn 1991 or Backus and Kehoe, forthcoming, for descriptions of other countries and time periods. 2 Although we concentrate mainly on these cyclical facts, we also mention an important long-run growth fact that is relevant for much of our discussion: total hours worked per capita do not display trend growth despite large secular increases in average productivity and real wages. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors LABOR MARKETS AND THE CYCLE 169 Gary D.Hansen, Randall Wright Real Business Cycle Theory preferences, households do not increase their work hours in response to permanent productivity growth. Thus, the nonseparable leisure model generates an increased standard deviation of hours worked relative to productivity without violating the long-run growth fact that hours worked per capita have not increased over long periods despite large increases in productivity. The second extension of the baseline real business cycle model we consider is to assume that labor is indivisible, so that workers can work either a fixed number of hours or not at all, as in Hansen 1985. In this version of the model, all variation in the labor input must come about by changes in the number of employed workers, which is the opposite of the standard model, where all variation comes about by changes in hours per worker. Although the data display variation along both margins, the indivisible labor model is perhaps a better abstraction, since the majority of the variance in the labor input in the United States can be attributed to changes in the number of employed workers. In the equilibrium of the indivisible labor model, individual workers are allocated to jobs randomly, and this turns out to imply that the aggregate economy displays a large labor supply elasticity even though individual hours do not respond at all to productivity or wage changes for continuously employed workers. The large aggregate labor supply elasticity leads to an increased standard deviation of hours relative to productivity, as compared to the baseline model. Neither nonseparable utility nor indivisible labor changes the result that the real business cycle model implies a large positive correlation between hours and productivity while the data display a near-zero correlation. This result arises because the model is driven by a single shock to the aggregate production function, which can be interpreted as shifting the labor demand curve along a stable labor supply curve and inducing a very tight positive relationship between hours and productivity. Hence, the next extension we consider is to introduce government spending shocks, as in Christiano and Eichenbaum 1992. If public consumption is an imperfect substitute for private consumption, then an increase in government spending has a negative wealth effect on individuals, which induces them to work more if leisure is a normal good. Therefore, government spending shocks can be interpreted as shifting the labor supply curve along the labor demand curve. Depending on the size of and the response to the two shocks, with this extension the model can generate a pattern of hours versus productivity closer to that found in the data. The final extension we consider is to introduce household production as in Benhabib, Rogerson, and Wright 1991. The basic idea is to recognize that agents derive utility from home-produced as well as marketproduced consumption goods and derive disutility from working in the home as well as in the market. In this version of the model, individuals, by working less at home, can increase hours of market work without reducing leisure as much. Therefore, the addition of household production increases the short-run labor supply elasticity and the standard deviation of hours relative to productivity. Furthermore, to the extent that shocks to household production are less than perfectly correlated with shocks to market production, individuals will have an incentive to substitute between home and market activity at a point in time. This is in addition to the standard incentive to substitute between market activity at different dates. Therefore, home production shocks, like government spending shocks, shift the labor supply curve and can generate a pattern of hours versus productivity closer to that found in the data. Our basic finding is that each of these four extensions to the baseline real business cycle model improves its performance quantitatively, even though the extensions work through very different economic channels. As will be seen, some of the resulting models seem to do better than others along certain dimensions, and some depend more sensitively than others on parameter values. Our goal here is not to suggest that one of these models is best for all purposes; which is best for any particular application will depend on the context. Rather, we simply want to illustrate here how incorporating certain natural features into the standard real business cycle model affects its ability to capture some key aspects of labor market behavior. The Facts In this section, we document the relevant business cycle facts. We consider several measures of hours worked and productivity and two sample periods (since some of the measures are available only for a shorter period). As in Prescott 1986, we define the business cycle as fluctuations around some slowly moving trend. For any given data series, we first take logarithms and then use the Hodrick-Prescott filter (as described in Prescott 1986) to remove the trend. Table 1 contains some summary statistics for quarterly U.S. data that are computed from deviations constructed in this manner. The sample period is from 1955:3 to 1988:2. The variables are y=output, c=consumption (nondurables plus services), i=fixed investment, h=hours worked, and w =average productivity (output divided by hours worked).4 For 3 Note that these preferences are nonseparable between leisure in different periods; they may or may not be separable between leisure and consumption in a given period. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 170 SOME EXTENSIONS Tables 1 and 2 Cyclical Properties of U.S. Time Series Table 2 1947:1–1991:3 * All series are quarterly, are in 1982 dollars, and have been logged and detrended with the Hodrick-Prescott filter. The output series, y, is the gross national product; c is consumption of nondurables and services; and i is fixed investment. Productivity is w=y/h. Sources: Citicorp’s Citibase data bank and Hansen 1991 4 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors LABOR MARKETS AND THE CYCLE 171 Gary D.Hansen, Randall Wright Real Business Cycle Theory each variable j, we report the following statistics: the (percent) standard deviation σj, the standard deviation relative to that of output σj /σy, and the correlation with output cor(j, y). We also report the relative standard deviation of hours to that of productivity σh/σw and the correlation between hours and productivity cor(h, w). We present statistics for four measures of h and w. Hours series 1 is total hours worked as recorded in the household survey and covers all industries. Hours series 2 is total hours worked as recorded in the establishment survey and covers only nonagricultural industries. These two hours series could differ for two reasons: they are from different sources, and they cover different industries.5 To facilitate comparison, we also report, in hours series 3, hours worked as recorded in the household survey but only for nonagricultural industries. Finally, hours series 4 is a measure of hours worked in efficiency units.6 The reason for the choice of 1955:3–1988:2 as the sample period is that hours series 3 and 4 are only available for this period. However, the other series are available for 1947:1–1991:3, and Table 2 reports statistics from this longer period for the available variables. Both Table 1 and Table 2 display the standard business cycle facts. All variables are positively correlated with output. Output is more variable than consumption and less variable than investment. Hours are slightly less variable than or about as variable as output, with σh/σy ranging between 0.78 and 1.01, depending on the hours series and the period. Overall, all variables are more volatile in the longer period, but the relative volatilities of the variables are about the same in the two periods. (An exception is investment, which looks somewhat less volatile relative to output in the longer period.) We want to emphasize two things. First, hours fluctuate more than productivity, with the magnitude of σh/σw ranging between 1.37 and 2.15, depending on the series and the period. Second, the correlation between hours and productivity is near zero or slightly negative, with cor(h, w) ranging between -0.35 and 0.10, depending on the series and the period. Chart 1 shows the scatter plot of h versus w from hours series 1 for the longer sample period (Plots from the other hours series look similar.) The Standard Model In this section, we present a standard real business cycle model and investigate its implications for the facts just described. The model has a large number of homogeneous households. The representative household has preferences defined over stochastic sequences of consumption ct and leisure lt, described by the utility function (1) where E denotes the expectation and β the discount factor, with β苸(0,1). The household has one unit of time each period to divide between leisure and hours of work: (2) The model has a representative firm with a constant returns-to-scale Cobb-Douglas production function that uses capital kt and labor hours ht to produce output yt: (3) where θ is the capital share parameter and is a stochastic term representing random technological progress. In general, we would assume that where is a constant yielding exogenous deterministic growth and zt evolves according to the process (4) where ρ苸(0,1) and εt is independent and normally distributed with mean zero and standard deviation σε. However, in this paper, we abstract from exogenous growth by setting 0.7 Capital evolves according to the law of motion (5) where δ is the depreciation rate and it investment. Finally, the economy must satisfy the resource constraint (6) We are interested in the competitive equilibrium of this economy. Since externalities or other distortions are not part of this model (or the other models that we consider), the competitive equilibrium is efficient. Hence, we can determine the equilibrium allocation by solving the social planner’s problem of maximizing the repre4 We use the letter w because average productivity is proportional to marginal productivity (given our functional forms), which equals the real wage rate in our models. 5 The establishment series is derived from payroll data and measures hours paid for, while the household series is taken from a survey of workers that attempts to measure hours actually worked. These two measures could differ, for example, because some workers may be on sick leave or vacation but still get paid. The household series is a better measure of the labor input, in principle, but because it is based on a survey of workers rather than payroll records, it is probably less accurate. 6 Efficiency units are constructed from hours series 3 by disaggregating individuals into age and sex groups and weighting the hours of each group by its relative hourly earnings; see Hansen 1991 for details. 7 Adding exogenous growth does not affect any of the statistics we report (as long as the parameters are recalibrated appropriately) given the way we filter the data; therefore, we set in order to simplify the presentation. See Hansen 1989. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 172 SOME EXTENSIONS sentative agent’s expected utility subject to feasibility constraints. That problem in this case is to maximize U subject to equations (2)–(6) and some initial conditions (k0, z0). The solution can be represented as a pair of stationary decision rules for hours and investment, ht=h*(kt, zt) and it=i*(kt, zt), that determine these two variables as functions of the current capital stock and technology shock. The other variables, such as consumption and output, can be determined from the decision rules using the constraints, while prices can be determined from the relevant marginal conditions. Standard numerical techniques are used to analyze the model. We choose functional forms and parameter values and substitute the constraint ct+it=f(zt, kt, ht) into the instantaneous return function u to reduce the problem to one of maximizing an objective function subject to linear constraints. Then we approximate the return function with a quadratic return function by taking a Taylor’s series expansion around the deterministic steady state. The resulting linear-quadratic problem can be easily solved for optimal linear decision rules, ht= h(kt, zt) and it=i(kt, zt); see Hansen and Prescott 1991 for details. Using these decision rules, we simulate the model, take logarithms of the artificially generated data, apply the Hodrick-Prescott filter, and compute statistics on the deviations (exactly as we did to the actual time series). We run 100 simulations of 179 periods (the number of quarters in our longer data set) and report the means of the statistics across these simulations. Preferences are specified so that the model is able to capture the long-run growth fact that per-capita hours worked display no trend despite large increases in productivity and real wages. When preferences are time separable, capturing this fact requires that the instantaneous utility function satisfy (7) or (8) where σ is a nonzero parameter and (l) is an increasing and concave function. (See King, Plosser, and Rebelo 1987, for example.) Intuitively, the growth facts imply that the wealth and substitution effects of long-run changes in productivity cancel, so the net effect is that hours worked do not change.8 We consider only preferences that satisfy (7) or (8); in fact, for convenience, we assume that (9) Parameter values are calibrated as follows. The discount factor is set to β=0.99 so as to imply a reasonable steady-state real interest rate of 1 percent per period (where a period is one quarter). The capital share parameter is set to θ=0.36 to match the average fraction of total income going to capital in the U.S. economy. The depreciation rate is set to δ=0.025, which (given the above-mentioned values for β and θ) implies a realistic steady-state ratio of capital to output of about 10 and a ratio of investment to output of 0.26. The parameter A in the utility function (9) is chosen so that the steady-state level of hours worked is exactly h=1/3, which matches the fraction of discretionary time spent in market work found in time-use studies (for example, Juster and Stafford 1991). Finally, the parameter ρ in (4) is set to ρ=0.95, and the standard deviation of ε is set to σε=0.007, which are approximately the values settled on by Prescott (1986). We focus on the following statistics generated by our artificial economy: the standard deviation of output; the standard deviations of consumption, investment, and hours relative to the standard deviation of output; the ratio of the standard deviation of hours to the standard deviation of productivity; and the correlation between hours and productivity. The results are shown in Table 3, along with the values for the same statistics from our longer sample from the U.S. economy (from Table 2). We emphasize the following discrepancies between the simulated and actual data. First, the model has a predicted standard deviation of output which is considerably less than the same statistic for the U.S. economy in either period. Second, the model predicts that σh/σw is less than one, while it is greater than one in the data. Third, the correlation between hours and productivity in the model is far too high. The result that output is not as volatile in the model economy as in the actual economy is not too surprising, since the model relies exclusively on a single technology shock, while the actual economy is likely to be subject to other sources of uncertainty as well. The result that in the model hours worked do not fluctuate enough relative to productivity reflects the fact that agents in the model are simply not sufficiently willing to substitute leisure in one period for leisure in 8 Other specifications can generate a greater short-run response of hours worked to productivity shocks; but while this is desirable from the point of view of explaining cyclical observations, it is inconsistent with the growth facts. For example, the utility function used in Greenwood, Hercowitz, and Huffman 1988, u(c, l)=(c+Al), has a zero wealth effect and hence a large labor supply elasticity, but implies that hours worked increase over time with productivity growth. This specification is consistent with balanced growth if we assume the parameter A grows at the same average rate as technology. Although such an assumption may seem contrived, it can be justified as the reduced form of a model with home production in which the home and market technologies advance at the same rate on average, as shown in Greenwood, Rogerson, and Wright 1992. 6 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors LABOR MARKETS AND THE CYCLE 173 Gary D.Hansen, Randall Wright Real Business Cycle Theory Table 3 Cyclical Properties of U.S. and Model-Generated Time Series * U.S. data here are the same as those in Table 2; they are for the longer time period: 1947:1–1991:3. ** The standard deviations and correlations computed from the models’ artificial data are the sample means of statistics computed for each of 100 simulations. Each simulation has 179 periods, the number of quarters in the U.S. data. Source: Citicorp’s Citibase data bank other periods. Finally, the result that hours and productivity are too highly correlated in the model reflects the fact that the only impulse driving the system is the aggregate technology shock. Chart 2 depicts the scatter plot between h and w generated by the model. Heuristically, Chart 2 displays a stable labor supply curve traced out by a labor demand curve shifting over time in response to technology shocks. This picture obviously differs from that in Chart 1. Nonseparable Leisure Following Kydland and Prescott (1982), we now attempt to incorporate the idea that instantaneous utility might depend not just on current leisure, but rather on a weighted average of current and past leisure. Hotz, Kydland, and Sedlacek (1988) find evidence in the panel data that this idea is empirically plausible. One interpretation they discuss concerns the fact that individuals need to spend time doing household chores, making repairs, and so on, but after doing so they can neglect these things for a while and spend more time working in the market until the results of their home work depreciate. The important impact of a nonseparable utility specification for our purposes is that, if leisure in one period is a relatively good substitute for leisure in nearby periods, then agents will be more willing to substitute intertemporally, and this increases the shortrun labor supply elasticity. Assume that the instantaneous utility function is u(ct, Lt)=log(ct)+Alog(Lt), where Lt is given by (10) and impose the restriction that the coefficients ai sum to one. If we also impose the restriction that (11) for i=1, 2, …, so that the contribution of past leisure to Lt decays geometrically at rate η, then the two parameters a 0 and η determine all of the coefficients in (10). Since Lt, and not simply lt, now provides utility, individuals are more willing to intertemporally substitute by working more in some periods and less in others. (At the same time, in a deterministic steady state or along a deterministic 7 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 174 SOME EXTENSIONS Charts 1–5 Hours Worked vs. Productivity in the Data and the Models 8 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors LABOR MARKETS AND THE CYCLE 175 Gary D.Hansen, Randall Wright Real Business Cycle Theory balanced growth path, this model delivers the correct prediction concerning the effect of productivity growth on hours worked.) The equilibrium can again be found as the solution to a social planner’s problem, which in this case maximizes U subject to (2)–(26), (10)–(11), and initial conditions. 9 The parameter values we use for the preference structure are a0=0.35 and η=0.10, which are the values implied by the estimates in Hotz, Kydland, and Sedlacek 1988; other parameter values are the same as in the preceding section. The results are in Table 3. Notice that output is more volatile here than in the standard model, with σy increasing from 1.30 to 1.51. Also, the standard deviation of hours worked relative to that of productivity has increased considerably, to σh/σw=1.63, and the correlation between hours and productivity has decreased somewhat to 0.80. Chart 3 depicts the scatter plot of h versus w generated by this model. Although these points trace out a labor supply curve that is flatter than the one in Chart 2, the model still does not generate the cloud in Chart 1. We conclude that introducing nonseparable leisure improves things in terms of σh/σw, but does little for cor(h, w). Indivisible Labor We now take up the indivisible labor model of Hansen (1985), in which individuals are constrained to work either zero or hˆ hours in each period, where 0<hˆ<1. Adding this constraint is meant to capture the idea that the production process has important nonconvexities or fixed costs that may make varying the number of employed workers more efficient than varying hours per worker. As originally shown by Rogerson (1984, 1988), in the equilibrium of this model, individuals will be randomly assigned to employment or unemployment each period, with consumption insurance against the possibility of unemployment. Thus, this model generates fluctuations in the number of employed workers over the cycle. As we shall see, it also has the feature that the elasticity of total hours worked increases relative to the standard model. Let πt be the probability that a given agent is ˆ employed in period t, so that Ht=πth is per-capita hours worked if we assume a large number of ex ante identical agents. Also, let c0t denote the consumption of an unemployed agent and c1t the consumption of an employed agent. As part of the dynamic social planning problem, πt, c0t, and c1t are chosen to maximize in each period, to (12) the following constraint: (13) where ct is total per-capita consumption. When u(c, l)=log(c) +Alog(l), the solution can be shown to imply that c1t=c0t= ct.10 Therefore, in the case under consideration, expected utility can be written (14) where B≡-Alog(1-hˆ)/hˆ>0 and, as defined above, Ht is hours worked per capita. Therefore, the indivisible labor model is equivalent to a divisible labor model with preferences described by (15) where Based on this equiva- lence, we can solve the indivisible labor model as if it were a divisible labor model with a different instantaneous utility function, by maximizing subject to (2)– (6) and initial conditions.11 Two features of the indivisible labor economy bear mention. First, as discussed earlier, fluctuations in the labor input come about by fluctuations in employment rather than fluctuations in hours per employed worker. This is the opposite of the standard model and is perhaps preferable, since the majority of the variance in total hours worked in the U.S. data is accounted for by variance in the number of workers. 12 Second, the indivisible labor model generates a large intertemporal substitution effect for the representative agent because instantaneous utility, is linear in H, and therefore the indifference curves between leisure in any two periods are linear. This is true despite the fact that hours worked are constant for a continuously employed worker. Return to Table 3 for the results of our simulations of this model. 13 The indivisible labor model is considerably more volatile than the standard model, with σy increasing from 1.30 to 1.73. Also, σh/σw has 9 For the solution techniques that we use, this problem is expressed as a dynamic program. The stock of accumulated past leisure is defined to be Xt, and we write Lt=a0lt+η(1-a0)X t Xt+1=(1-η)X t+l t. These equations replace (10) and (11) in the recursive formulation. 10 This implication follows from the fact that u is separable in c and l and does not hold for general utility functions; see Rogerson and Wright 1988. 11 Since the solution to the planner’s problem in the indivisible labor model involves random employment, we need to use some type of lottery or sunspot equilibrium concept to support it as a decentralized equilibrium; see Shell and Wright, forthcoming. 12 See Hansen 1985 for the U.S. data. Note, however, that European data display greater variance in hours per worker than in the number of workers; see Wright 1991, p. 17. 13 The new parameter B is calibrated so that steady-state hours are again equal to 1/3; the other parameters are the same as in the standard model. 9 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 176 SOME EXTENSIONS increased from 0.94 to 2.63, actually somewhat high when compared to the U.S. data. Of course, this model is extreme in the sense that all fluctuations in the labor input result from changes in the number of employed workers, and models in which both the number of employed workers and the number of hours per worker vary fall somewhere between the standard divisible labor model and the indivisible labor model with respect to this statistic. (See Kydland and Prescott 1991 or Cho and Cooley 1989, for example.) Finally, the model implies that cor(h, w)=0.76, slightly lower than the models discussed above but still too high. For the sake of brevity, the scatter plot between h and w is omitted; for the record, it looks similar to the one in Chart 3, although the indivisible labor model displays a little more variation in hours worked. Government Spending We now introduce stochastic government spending, as in Christiano and Eichenbaum 1992. (That paper also provides motivation and references to related work.) Assume that government spending, gt, is governed by (16) respectively. (In addition, the average of gt/yt in our sample, which is 0.22, is used to calibrate ) For the results, turn again to Table 3. The government spending model actually behaves very much like the standard model, except that the correlation between hours and productivity decreases to cor(h, w)=0.49, which is better than the previous models although still somewhat larger than the U.S. data. Chart 4 displays the scatter plot generated by the model with only government spending shocks (that is, with the variance in the technology shock set to σε=0), and Chart 5 displays the scatter plot for the model with both shocks. These charts illustrate the intuition behind the results: technology shocks shift labor demand and trace out the labor supply curve, government shocks shift labor supply and trace out the labor demand curve, and both shocks together generate a combination of these two effects. The net results will be somewhat sensitive to the size of and the response to the two shocks; however, for the estimated parameter values, this model generates a scatter plot that is closer to the data than does the standard model.15 Home Production where ⑀(0, 1) and µt is independent and normally distributed with mean zero and standard deviation σµ. Furthermore, as in Christiano and Eichenbaum 1992, assume that µt is independent of the technology shock. Also assume that government spending is financed by lump-sum taxation and that it enters neither the utility function nor the production function.14 Then the equilibrium allocation for the model can be found by solving the planner’s problem of maximizing U subject to (16), (2)–(5), and, instead of (6), the new resource constraint (17) An increase in gt is a pure drain on output here. Since leisure is a normal good, the negative wealth effect of an increase in gt induces households to work more. Intuitively, shocks to gt shift the labor supply curve along the demand curve at the same time that technology shocks shift the labor demand curve along the supply curve. This first effect produces a negative relationship between hours and productivity, while the second effect produces a positive relationship. The net effect on the correlation between hours and productivity in the model depends on the size of the gt shocks and on the implied wealth effect, which depends, among other things, on the parameter λ in the law of motion for g t (because temporary shocks have a smaller wealth effect than permanent shocks). Hence, the calibration of this law of motion is critical. An ordinary least squares regression based on equation (16) yields estimates for λ and σµ of 0.96 and 0.021, We now consider the household production model analyzed in Benhabib, Rogerson, and Wright 1991. (That paper also provides motivation and references to related work.) Instantaneous utility is still written u(c, l)=log(c)+Alog(l), but now consumption and leisure have a different interpretation. We assume that (18) (19) 14 A generalization is to assume that instantaneous utility can be written u(C, l), where C=C(c, g) depends on private consumption and government spending. The special case where C=c is the one we consider here, while the case where C=c+g can be interpreted as the standard model, since then increases in g can be exactly offset by reductions in c and the other variables will not change. Therefore, the model with C= c+g generates exactly the same values of all variables, except that c+g replaces c. The assumption that c and g are perfect substitutes implies that they are perfectly negatively correlated, however. A potentially interesting generalization would be to assume that C(c, g)=[αcφ+(1-α)gφ]1/φ where l/(l-φ) is the elasticity of substitution. 15 The size of the wealth effect depends on the extent to which public consumption and private consumption are substitutes. For example, if they were perfect substitutes, then a unit increase in g would simply crowd out a unit of c with no effect on hours worked or any of the other endogenous variables. We follow Christiano and Eichenbaum 1992 in considering the extreme case where g does not enter utility at all. Also, the results depend on the (counterfactual) assumption that the shocks to government spending and technology are statistically independent. Finally, the results depend on the estimates of the parameters in the law of motion (16). The estimates in the text are from the period 1947:1–1991:3 and are close to the values used in Christiano and Eichenbaum 1992. Estimates from our shorter sample period, 1955:3–1988:2, imply a higher λ of 0.98 and a lower σµ of 0.012, which in simulations yield cor(h, w)=0.65. 10 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors LABOR MARKETS AND THE CYCLE Gary D.Hansen, Randall Wright Real Business Cycle Theory where cMt is consumption of a market-produced good, cHt is consumption of a home-produced good, hMt is hours worked in the market sector, and hHt is hours worked in the home, all in period t. Notice that the two types of work are assumed to be perfect substitutes, while the two consumption goods are combined by an aggregator that implies a constant elasticity of substitution equal to 1/(1-e). This model has two technologies, one for market production and one for home production: (20) (21) where and are the capital share parameters. The two technology shocks follow the processes (22) 177 σh/σw=1.92. And cor(h, w) has decreased to 0.49, the same as in the model with government spending.17 The intuition behind these results is that agents substitute in and out of market activity more in the home production model than in the standard model because they can use nonmarket activity as a buffer. The degree to which agents do this depends on their willingness to substitute cM for cH, as measured by e, and on their incentive to move production between the two sectors, as measured by γ. (Lower values of γ entail more frequent divergence between zM and zH and, hence, more frequent opportunities to specialize over time.) Note that some aspects of the results do not actually depend on home production being stochastic.18 However, the correlation between productivity and market hours does depend critically on the size of the home technology shock, exactly as it depends on the size of the second shock in the government spending model. We omit the home production model’s scatter plot between h and w, but it looks similar to that of the model with government shocks. Conclusion (23) where the two innovations are normally distributed with standard deviations σM and σH, have a contemporaneous correlation γ=cor(εMt, εHt), and are independent over time. In each period, a capital constraint holds: kMt+kHt=kt, where total capital evolves according to kt+1=(1-δ)kt+it. Finally, the constraints imply that all (24) (25) new capital is produced in the market sector. The parameters β, , ␦, and are set to the values used in the previous sections. The two utility parameters A and a are set to deliver steady-state values of hM=0.33 and hH=0.28, as found in the timeuse studies (Juster and Stafford 1991), and the capital share parameter in the household sector is set to =0.08, implying a steady-state ratio of c H/c M of approximately 1/4.16 The variances of the two shocks are assumed to be the same: H=M =0.007. The parameter e, which determines the elasticity of substitution between cM and cH, and γ, which is the correlation between ε M and ε H , are set to the benchmark values used in Benhabib, Rogerson, and Wright 1991: e=0.8 and γ=2/3. The results are at the bottom of Table 3. In the home production model, output is more volatile than in the standard model and about as volatile as in the indivisible labor model. The standard deviation of hours relative to productivity has increased considerably compared to the standard model, to We have presented several extensions to the standard real business cycle model and analyzed the extent to which they help account for the U.S. business cycle facts, especially those facts concerning hours and productivity. Introducing nonseparable leisure, indivisible labor, or home production increases the elasticity of hours worked with respect to short-run productivity changes. Introducing a second shock, either to government spending or to the home production function, reduces the correlation between hours worked and productivity.19 Note that our goal has not been to convince you that any of these models is unequivocally to be preferred. Our goal has been simply to explain some commonly used real business cycle models and compare their implications for the basic labor market facts. 16 The two parameters and can be calibrated to match the observed average levels of market capital (producer durables and nonresidential structures) and home capital (consumer durables and residential structures) in the U.S. economy. This requires a lower value for and a higher value for than used here, as discussed in Greenwood, Rogerson, and Wright 1992. 17 The exact results are somewhat sensitive to changes in the parameters e and ␥, for reasons discussed in the next paragraph. 18 Even if the variance of the shock to the home technology is set to zero, shocks to the market technology will still induce relative productivity differentials across sectors. And even if the two shocks are perfectly correlated and of the same magnitude, agents will still have an incentive to switch between sectors over time because capital is produced exclusively in the market. It is these effects that are behind the increase in the labor supply elasticity. 19 Other extensions not considered here can also affect the implications of the model for the labor market facts, including distorting taxation as in Braun 1990 or McGrattan 1991 and nominal contracting as in Cho and Cooley 1990. 11 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 178 SOME EXTENSIONS References Backus, David K., and Kehoe, Patrick J.Forthcoming. International evidence on the historical properties of business cycles. American Economic Review. Benhabib, Jess; Rogerson, Richard; and Wright, Randall. 1991. Homework in macroeconomics: Household production and aggregate fluctuations. Journal of Political Economy 99 (December): 1166–87. Blackburn, Keith, and Ravn, Morten O. 1991. Contemporary macroeconomic fluctuations: An international perspective. Memo 1991–12. University of Aarhus Center for International Economics. Braun, R.Anton. 1990. The dynamic interaction of distortionary taxes and aggregate variables in postwar U.S. data. Working Paper. University of Virginia. Cho, Jang-Ok, and Cooley, Thomas F. 1989. Employment and hours over the business cycle. Working Paper 132. Rochester Center for Economic Research. University of Rochester. _______. 1990. The business cycle with nominal contracts. Working Paper 260. Rochester Center for Economic Research. University of Rochester. Christiano, Lawrence J., and Eichenbaum, Martin. 1992. Current realbusiness-cycle theories and aggregate labor-market fluctuations. American Economic Review 82 (June): 430–50. Greenwood, Jeremy; Hercowitz, Zvi; and Huffman, Gregory W. 1988. Investment, capacity utilization and the real business cycle. American Economic Review 78 (June): 402–17. Greenwood, Jeremy; Rogerson, Richard; and Wright, Randall. 1992. Household production in real business cycle theory. Manuscript. University of Western Ontario. Hansen, Gary D. 1985. Indivisible labor and the business cycle. Journal of Monetary Economics 16 (November): 309–27. _______. 1989. Technical progress and aggregate fluctuations. Department of Economics Working Paper 546. University of California, Los Angeles. _______. 1991. The cyclical and secular behavior of the labor input: Comparing efficiency units and hours worked. Manuscript. University of California, Los Angeles. Hansen, Gary D., and Prescott, Edward C. 1991. Recursive methods for computing equilibria of business cycle models. Discussion Paper 36. Institute for Empirical Macroeconomics (Federal Reserve Bank of Minneapolis). Hotz, V.Joseph; Kydland, Finn E.; and Sedlacek, Guilherme L. 1988. Intertemporal preferences and labor supply. Econometrica 56 (March): 335–60. Juster, F.Thomas, and Stafford, Frank P. 1991. The allocation of time: Empirical findings, behavioral models, and problems of measurement. Journal of Economic Literature 29 (June): 471–522. King, Robert G.; Plosser, Charles I.; and Rebelo, Sergio T. 1987. Production, growth and cycles: Technical appendix . Manuscript. University of Rochester. Kydland, Finn E., and Prescott, Edward C. 1982. Time to build and aggregate fluctuations. Econometrica 50 (November): 1345–70. _______. 1990. Business cycles: Real facts and a monetary myth. Federal Reserve Bank of Minneapolis Quarterly Review 14 (Spring): 3–18. _______. 1991. Hours and employment variation in business cycle theory. Economic Theory 1:63–81. McGrattan, Ellen R. 1991. The macroeconomic effects of distortionary taxation. Discussion Paper 37. Institute for Empirical Macroeconomics (Federal Reserve Bank of Minneapolis). Prescott, Edward C. 1986. Theory ahead of business cycle measurement Federal Reserve Bank of Minneapolis Quarterly Review 10 (Fall): 9–22. Rogerson, Richard. 1984. Topics in the theory of labor markets. Ph.D. dissertation. University of Minnesota. _______. 1988. Indivisible labor, lotteries and equilibrium. Journal of Monetary Economics 21 (January): 3–16. Rogerson, Richard, and Wright, Randall. 1988. Involuntary unemployment in economies with efficient risk sharing. Journal of Monetary Economics 22 (November): 501–15. Shell, Karl, and Wright, Randall. Forthcoming. Indivisibilities, lotteries and sunspot equilibria. Economic Theory. Wright, Randall. 1991. The labor market implications of unemployment insurance and short-time compensation. Federal Reserve Bank of Minneapolis Quarterly Review 15 (Summer): 11–19. 12 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors CHAPTER 10 179 Current Real-Buisness-Cycle Theories and Aggregate Labor-Market Fluctuations By LAWRENCE J.CHRISTIANO AND MARTIN EICHENBAUM* Hours worked and the return to working are weakly correlated. Traditionally, the ability to account for this fact has been a litmus test for macroeconomic models. Existing realbusiness-cycle models fail this test dramatically. We modify prototypical real-business-cycle models by allowing government consumption shocks to influence labor-market dynamics. This modification can, in principle, bring the models into closer conformity with the data. Our empirical results indicate that it does. (JEL E32, C12, C52, C13, C51) In this paper, we assess the quantitative implications of existing real-business-cycle (RBC) models for the time-series properties of average productivity and hours worked. We find that the single most salient shortcoming of existing RBC models lies in their predictions for the correlation between these variables. Existing RBC models predict that this correlation is well in excess of 0.9, whereas the actual correlation is much closer to zero.1 This shortcoming leads us to add to the RBC framework aggregate demand shocks that arise from stochastic movements in government consumption. According to our empirical results, this change substantially improves the models’ empirical performance. The ability to account for the observed correlation between the return to working and the number of hours worked has traditionally been a litmus test for aggregate economic models. Thomas J.Sargent (1987 p. 468), for example, states that one of the primary empirical patterns casting doubt on the classical and Keynesian models has been the observation by John T.Dunlop (1938) and Lorie Tarshis (1939) “alleging the failure of real wages to move countercyclically.” The classical and Keynesian models share the assumption that real wages and hours worked lie on a stable, downward-sloped marginal productivity-of-labor curve. 2 Consequently, they both counterfactually predict a strong negative correlation between real wages and hours worked. Modern versions of what Sargent (1987 p. 468) calls the “Dunlop-Tarshis observation” continue to play a central role in assessing the empirical plausibility of different business-cycle models.3 In discussing Stanley Fischer’s (1977) sticky-wage business-cycle *Christiano: Federal Reserve Bank of Minneapolis, Minneapolis, MN 55480; Eichenbaum: Northwestern University, Evanston, IL 60208, National Bureau of Economic Research, and Federal Reserve Bank of Chicago. This paper is a substantially revised version of NBER Working Paper No. 2700, “Is Theory Really Ahead of Measurement? Current Real Business Cycle Theories and Aggregate Labor Market Fluctuations.” We thank Rao Aiyagari, Paul Gomme, Finn Kydland, Ed Prescott, and Mark Watson for helpful conversations. Any views expressed here are ours and not necessarily those of any part of the Federal Reserve System. 1 This finding is closely related to Bennett McCallum’s (1989) observation that existing RBC models generate grossly counterfactual predictions for the correlation between average productivity and output. 2 As John Maynard Keynes (1935 p. 17) says, “…I am not disputing this vital fact which the classical economists have (rightly) asserted as indefeasible. In a given state of organisation, equipment and technique, the real wage earned by a unit of labour has a unique (inverse) correlation with the volume of employment.” 3 For example, Robert J.Barro and Herschel I.Grossman (1971) cite the Dunlop-Tarshis observation to motivate their work on disequilibrium theories. Also, Edmund S.Phelps and Sidney G.Winter, Jr. (1970 p. 310) and Franco Modigliani (1977 p. 7) use this observation to motivate their work on noncompetitive approaches to macroeconomics. Finally, Robert E.Lucas, Jr. (1981 p. 13) cites the Dunlop-Tarshis observation in motivating his work on capacity and overtime. 430 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 180 SOME EXTENSIONS VOL. 82 NO. 3 CHRISTIANO AND EICHENBAUM: CURRENT RBC THEORIES model, for example, Bennett McCallum (1989 p. 191) states that …the main trouble with the Fischer model concerns its real wage behavior. In particular, to the extent that the model itself explains fluctuations in output and empløyment, these should be inversely related to real wage movements: output should be high, according to the model, when real wages are low. But in the actual U.S. economy there is no strong empirical relation of that type. In remarks particularly relevant to RBC models, Robert E.Lucas (1981 p. 226) says that “observed real wages are not constant over the cycle, but neither do they exhibit consistent pro- or countercyclical tendencies. This suggests that any attempt to assign systematic real wage movements a central role in an explanation of business cycles is doomed to failure.” Existing RBC models fall prey to this (less wellknown) Lucas critique. Unlike the classical and Keynesian models, which understate the correlation between hours worked and the return to working, existing RBC models grossly overstate that correlation. According to existing RBC models, the only impulses generating fluctuations in aggregate employment are stochastic shifts in the marginal product of labor. Loosely speaking, the time series on hours worked and the return to working are modeled as the intersection of a stochastic labor-demand curve 4 Although Finn E.Kydland and Edward C.Prescott (1982) and Prescott (1986) never explicitly examine the hours/real-wage correlation implication of existing RBC models, Prescott (1986 p. 21) does implicitly acknowledge that the failure to account for the Dunlop-Tarshis observation is the key remaining deviation between economic theory and observations: “The key deviation is that the empirical labor elasticity of output is less than predicted by theory.” To see the connections, denote the empirical 431 with a fixed labor-supply curve. Not surprisingly, therefore, these theories predict a strong positive correlation between hours worked and the return to working.4 Several strategies exist for modeling the observed weak correlation between measures of these variables. One is to consider models in which the return to working is unaffected by shocks to agents’ environments, regardless of whether the shocks are to aggregate demand or to aggregate supply. Pursuing this strategy, Olivier Jean Blanchard and Stanley Fischer (1989 p. 372) argue that the key assumption of Keynesian macro models—nominal wage and price stickiness—is motivated by the view that aggregate demand shocks affect employment but not real wages. Another strategy is simply to abandon one-shock models of aggregate fluctuations and suppose that the business cycle is generated by a variety of impulses. Under these conditions, the Dunlop-Tarshis observation imposes no restrictions per se on the response of real wages to any particular type of shock. Given a specific structural model, however, it does impose restrictions on the relative frequency of different types of shocks. This suggests that one strategy for reconciling existing RBC models with the Dunlop-Tarshis observation is to find measurable economic impulses that shift the labor-supply function.5 With different impulses shifting the laborsupply and labor-demand functions, there is no a priori reason for hours worked to be labor elasticity by . By definition, ≡corr(y, n)σy/σn, where corr(i, j) is the correlation between i and j, σi is the standard deviation of i, y is the logarithm of detrended output, and n is the logarithm of hours. Simple arithmetic yields corr(y-n, n)=[η-1](n/y-n). If, as Prescott claims, RBC models do well at reproducing the empirical estimates of n/y-n, then saying that the models overstate is equivalent to saying that they overstate corr(y-n, n). In Prescott’s model and with his assumed market structure, corr(y-n, n) is exactly the same as the correlation between real wages and hours worked. (Also, under log detrending, y-n is log detrended productivity.) 5 An alternative strategy is pursued by Valerie R.Bencivenga (1992), who allows for shocks to labor suppliers’ preferences. Matthew D.Shapiro and Mark W.Watson (1988) also allow for unobservable shocks to the labor-supply function. Jess Benhabib et al. (1991) and Jeremy Greenwood and Zvi Hercowitz (1991) explore the role of shocks to the home production technology. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors AGGREGATE LABOR MARKET FLUCTUATIONS 432 THE AMERICAN ECONOMIC REVIEW JUNE correlated in any particular way with the return to working. Candidates for such shocks include tax rate changes, innovations to the money supply, demographic changes in the labor force, and shocks to government spending. We focus on the last of these. By ruling out any role for government-consumption shocks in labormarket dynamics, existing RBC models implicitly assume that public and private consumption have the same impact on the marginal utility of private spending. Robert J.Barro (1981, 1987) and David Alan Aschauer (1985) argue that if $1 of additional public consumption drives down the marginal utility of private consumption by less than does $1 of additional private consumption, then positive shocks to government consumption in effect shift the labor-supply curve outward. With diminishing labor productivity, but without technology shocks, such impulses will generate a negative correlation between hours worked and the return to working in RBC models. In our empirical work, we measure the return to working by the average productivity of labor rather than real wages. We do this for both empirical and theoretical reasons. From an empirical point of view, our results are not very sensitive to whether the return to working is measured by real wages or average productivity: Neither displays a strong positive correlation with hours worked, so it seems appropriate to refer to the low correlation between the return to working and hours worked as the Dunlop-Tarshis observation, regardless of whether the return to working is measured by the real wage or average productivity. From a theoretical point of view, a variety of ways exist to support the quantity allocations emerging from RBC models. By using average productivity as our measure of the return to working, we avoid imposing the assumption that the market structure is one in which real wages are equated to the marginal product of labor on a period-by-period basis. Also, existing parameterizations of RBC models imply that marginal and average productivity of labor are proportional to each 181 1992 other. For the calculations we perform, the two are interchangeable. Our empirical results show that incorporating government into the analysis substantially improves the RBC models’ performance. Interestingly, the impact of this perturbation is about as large as allowing for nonconvexities in labor supply of the type stressed by Gary D.Hansen (1985) and Richard Rogerson (1988). Once government is incorporated into the analysis, we cannot reject the hypothesis that a version of the Hansen-Rogerson indivisible-labor model is consistent with both the observed correlation between hours worked and average productivity and the observed volatility of hours worked relative to average productivity. This is not true if government is excluded from the analysis. The paper is organized as follows. In Section I, we describe a general equilibrium model that nests as special cases a variety of existing RBC models. In Section II, we present our econometric methodology for estimating and evaluating the empirical performance of the model. In Section III, we present our empirical results. In Section IV, we offer some concluding remarks. I. Two Prototypical Real-Business-Cycle Models In this section, we present two prototypical real-business-cycle models. One is a stochastic version of the one-sector growth model considered by Kydland and Prescott (1980 p. 174). The other is a version of the model economy considered by Hansen (1985) in which labor supply is indivisible. In both of our models, we relax the assumption implicit in existing RBC models that public and private spending have identical effects on the marginal utility of private consumption. We make the standard RBC assumption that the time series on the beginning-of-periodt per capita stock of capital (kt), private time-t and hours worked at time t consumption (nt) correspond to the solution of a socialplanning problem which can be decentralized as a Pareto-optimal competitive equilibrium. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 182 SOME EXTENSIONS VOL. 82 NO. 3 CHRISTIANO AND EICHENBAUM: CURRENT RBC THEORIES The following problem nests both our models as special cases. Let N be a positive scalar that denotes the time-t endowment of the representative consumer, and let γ be a positive scalar. The social planner ranks streams of consumption services (ct), leisure (N-nt), and publicly provided goods and services (gt) according to the criterion function (1) Following Barro (1981, 1987), Roger C. Kormendi (1983), and Aschauer (1985), we suppose that consumption services are related to private and public consumption as follows: (2) where α is a parameter that governs the sign and magnitude of the derivative of the with respect to gt.6 marginal utility of Throughout, we assume that agents view gt as an uncontrollable stochastic process. In addition, we suppose that gt does not depend on the current or past values of the endogenous variables in the model.7 We consider two specifications for the function V(·). In the divisible-labor model, V(·) is given by (3) 433 (4) This specification can be interpreted in at least two ways. One is that the specification simply reflects the assumption that individual utility functions are linear in leisure. The other interpretation builds on the assumption that labor supply is indivisible. Under this second interpretation, individuals can either work some positive number of hours or not work at all. Assuming that agents’ utility functions are separable across consumption and leisure, Rogerson (1988) shows that a market structure in which individuals choose lotteries rather than hours worked will support a Paretooptimal allocation of consumption and leisure. The lottery determines whether individuals work or not. Under this interpretation, (4) represents a reduced-form preference-ordering that can be used to derive the Pareto-optimal allocation by solving a fictitious social-planning problem. This is the specification used by Hansen (1985). Per capita output yt is produced using the Cobb-Douglas production function (5) where 0<<1 and zt is an aggregate shock to technology that has the time-series representation (6) In the indivisible-labor model, V(·) is given by Here t is a serially uncorrelated independent and identically distributed process with mean and standard error . The aggregate resource constraint is given by (7) 6 We can generalize the criterion function (1) by writing it as ln(ct)+␥V(N-nt)+φ (gt), where φ (·) is some positive concave function. As long as gt is modeled as an exogenous stochastic process, the presence of such a term has no impact on the competitive equilibrium. However, the presence of φ (gt)>0 means that agents do not necessarily feel worse off when gt is increased. The fact that we have set φ (·)≡0 reflects our desire to minimize notation, not the view that the optimal level of gt is zero. 7 Under this assumption, gt is isomorphic to an exogenous shock to preferences and endowments. Consequently, existing theorems which establish that the competitive equilibrium and the social-planning problem coincide are applicable. That is, per capita consumption and investment cannot exceed per capita output. At time 0, the social planner chooses to contingency plans for maximize (1) subject to (3) or (4), (5)–(7), k0, and a law of motion for gt. Because of the nonsatiation assumption implicit in (1), we can, without loss of generality, impose strict equality © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors AGGREGATE LABOR MARKET FLUCTUATIONS 434 THE AMERICAN ECONOMIC REVIEW JUNE 183 1992 in (7). Substituting (2), (5), and this version of (7) into (1), we obtain the following social-planning problem: maximize subject to k0, a law of motion same as those that permanently enhance the economy’s productive ability. Substituting (9) into (8), we obtain the criterion function: (8) (11) where (12) for gt, and V(·) given by either (3) or (4), by choice of contingency plans for {kt+1, nt: t≥0}. It is convenient to represent the socialplanning problem (8) in a way such that all of the planner’s decision variables converge in nonstochastic steady state. To that end, we define the following detrended variables: (9) To complete our specification of agents’ environment, we assume that g-t evolves according (10) to where is the mean of and µt is the innovation in with standard deviation µ. Notice that gt has two components, zt and . Movements in zt produce permanent changes in the level of government consumption, whereas movements in g- t produce temporary changes in gt. With this specification, the factors giving rise to permanent shifts in government consumption are the and where and V(·) is given by either (3) or (4). Consequently, the original planning problem is equivalent to the prob(10), and lem of maximizing (11), subject to (12), and V(·) is given by either (3) or (4). Since is beyond the planner’s control, it can be disregarded in solving the planner’s problem. The only case in which an analytical solution for this problem is possible occurs when ␣=␦=1 and the function V( ·) is given by (3). John B.Long, Jr., and Charles I.Plosser (1983) provide one analysis of this case. Analytical solutions are not available for general values of ␣ and ␦. We use Christiano’s (1988) log-linear modification of the procedure used by Kydland and Prescott (1982) to obtain an approximate solution to our social-planning problem. In particular, we approximate the optimal decision rules with the solution to the linear-quadratic problem obtained when the function r in (12) is replaced by a function R, which is quadratic in ln(nt), and t. The function R is the secondorder Taylor expansion of r[exp(A1), exp(A2), exp(A3), exp(A4), A5] about the point © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 184 SOME EXTENSIONS VOL. 82 NO. 3 CHRISTIANO AND EICHENBAUM: CURRENT RBC THEORIES Here n and denote the steady-state values of in the nonstochastic version of (11) nt and obtained by setting λ=µ=0. Results in Christiano (1988) establish that the decision rules which solve this problem are of the form (13) and (14) In (13) and (14), rk, dk, ek, rn, dn, and en are scalar functions of the model’s underlying structural parameters.8 To gain intuition for the role of in aggregate labor-market fluctuations, it is useful to discuss the impact of three key parameters (␣, , and ␥) on the equilibrium response of nt to This response is governed by the coefficient dn. First, notice that when ␣=1 the only way and gt enter into the social planner’s preferences and constraints is through their sum, Thus, exogenous shocks to gt induce one-forleaving other one offsetting shocks in variables like yt, kt+1, and nt unaffected. This implies that the coefficients dn and dk in the planner’s decision rules for kt+1 and nt both equal zero. Consequently, the absence of a role for gt in existing RBC models can be interpreted as reflecting the assumption that ␣=1. Second, consider what happens when ␣<1. The limiting case of ␣=0 is particularly useful for gaining intuition. Government consumption now is formally equivalent to a pure resource drain on the economy; agents respond to an increase in government consumption as if they had suffered a reduction in their wealth. (As footnote 6 indicates, this does not imply that they have 8 Christiano (1987a, 1988 footnotes 9, 18) discusses the different properties of the log-linear approximation that we use here and linear approximations of the sort used by Kydland and Prescott (1982). 435 suffered a reduction in utility.) The coefficient dn is positive, since we assume that leisure is a are normal good. That is, increases in associated with increases in nt and decreases in y t/n t . Continuity suggests that d n is decreasing in a. The same logic suggests that dn is an increasing function of , since the wealth effect of a given shock to is increasing in . For a formal analysis of the effects of government consumption in a more general environment than the one considered here, see S.Rao Aiyagari et al. (1990). Finally, consider the impact of ␥ on aggregate labor-market fluctuations. In several experiments, we found that en and dn were increasing in ␥ (for details, see Christiano and Eichenbaum [1990a]). To gain intuition into this result, think of a version of the divisible-labor model in which the gross investment decision rule is fixed exogenously. In this simpler model economy, labor-market equilibrium is the result of the intersection of static labor-supply and labordemand curves. Given our assumptions regarding the utility function, the response of labor supply to a change in the return to working is an increasing function of ␥ ; that is, the labor-supply curve becomes flatter as ␥ increases. By itself, this makes the equilibrium response of nt to t (which shifts the labor-demand curve) an increasing function of ␥. This relationship is consistent with the finding that en is increasing in ␥ in our model. With respect to d n , it is straightforward to show that, in the static framework, the extent of the shift in the labor-supply curve induced by a change in is also an increasing function of ␥ . This is also consistent with the finding that dn is an increasing function of ␥ in our model. That en and dn are increasing in ␥ leads us to expect that the volatility of hours worked will also be an increasing function of ␥ . However, we cannot say a priori what impact larger values of ␥ will have on the DunlopTarshis correlation, because larger values of en drive that correlation up, but larger values of dn drive it down. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors AGGREGATE LABOR MARKET FLUCTUATIONS 436 THE AMERICAN ECONOMIC REVIEW JUNE II. Econometric Methodology In this section, we describe three things: our strategy for estimating the structural parameters of the model and various second moments of the data, our method for evaluating the model’s implications for aggregate labor-market fluctuations, and the data used in our empirical analysis. While similar in spirit, our empirical methodology is quite different from the methods typically used to evaluate RBC models. Much of the existing RBC literature makes little use of formal econometric methods, either when model parameter values are selected or when the fully parameterized model is compared with the data. Instead, the RBC literature tends to use a variety of informal techniques, often referred to as calibration. In contrast, we use a version of Lars Peter Hansen’s (1982) generalized method-of-moments (G M M) procedure at both stages of the analysis. Our estimation criterion is set up so that, in effect, estimated parameter values equate model and sample first moments of the data. It turns out that these values are very similar to the values used in existing RBC studies. An important advantage of our GMM procedures, however, is that they let us quantify the degree of uncertainty in our estimates of the model’s parameters. This turns out to be an important ingredient of our model-evaluation techniques. A. Estimation Now we will describe our estimation strategy. The parameters of interest can be divided into two groups. Let ⌿ 1 denote the model’s eight structural parameters: (15) The parameters N, , and a were not estimated. Instead, we fixed N at 1,369 hours per quarter and set the parameter  so as to imply a 3percent annual subjective discount rate; that is, =(1.03)-0.25. Two alternative values of α were considered: α=0 and α=1. and Given estimated values of ⌿ 1, 185 1992 distribution assumptions on µt and t, our model provides a complete description of the data-generating process. (Here T denotes the number of observations in our sample.) This can be used to compute the second moments of all the variables of the model. Suppose, for the time being, that we can abstract from say, because we sampling uncertainty in have a large data sample. Then the second will coincide with moments implied by the second moments of the stochastic process generating the data only if the model has been specified correctly. This observation motivates our strategy for assessing the empirical plausibility of the model. First we calculate selected second moments of the data using our model . Then we estimate the same evaluated at second moments directly, without using the model. Our test then compares these two sets of second moments and determines whether the differences between them can be accounted for by sampling variation under the null hypothesis that the model is correctly specified. To this end, it is useful to define ⌿ 2 to be various second moments of the data. Our dkt, kt, yt, (y/n)t and gt all display measures of marked trends, so some stationarity-inducing transformation of the data must be adopted for second moments to be well defined. (Here dk t denotes gross investment.) The transformation we used corresponds to the Hodrick and Prescott (H P) detrending procedure discussed by Robert J.Hodrick and Prescott (1980) and Prescott (1986). We used the H P transformation because many researchers, especially Kydland and Prescott (1982, 1988), G.Hansen (1985), and Prescott (1986), have used it to investigate RBC models. Also, according to our model, the logarithms dkt, kt, yt, (y/n)t, and gt are all differenceof stationary stochastic processes. That the HP filter is a stationarity-inducing transformation for such processes follows directly from results of Robert G.King and Sergio T.Rebelo (1988). We also used the first-difference filter in our analysis. Since the results are not substantially different from those reported © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 186 SOME EXTENSIONS VOL. 82 NO. 3 CHRISTIANO AND EICHENBAUM: CURRENT RBC THEORIES 437 here, we refer the reader to Christiano and Eichenbaum (1990a) for details. The parameters in ⌿ 2 are rate of substitution of goods in consumption equals the time-t expected value of the marginal return to physical investment in capital. Therefore, (16) (18) where σx denotes the standard deviation of the variable x, with x ={cp, y, dk, n, y/n, g}, and corr(y/ n, n) denotes the correlation between y/n and n. 1. The Unconditional Moments Underlying Our Estimator of ⌿ 1.—The procedure we used to estimate the elements of ⌿ 1 can be described as follows. Our estimator of ␦ is, roughly, the rate of depreciation of capital implicit in the empirical capital-stock and investment series. The estimators of and ␥ are designed to allow the model to reproduce the average value of the capital: output ratio and hours worked observed in the data. The point estimates of , and µ are obtained by applying ordinary least squares to data on g t/z t, where z t is constructed using the estimated value of . Finally, our point estimates of and are the mean growth rate of output and the standard deviation of the growth rate of zt, respectively. We map these estimators into a G M M framework to get an estimate of the sampling distribution of our estimator of ⌿ 1. We need that estimate for our diagnostic procedures. To use GMM, we express the estimator of ⌿ 1, as the solution to the sample analog of first-moment conditions. We now describe these conditions. According to our model, ␦ =1+(dk t/k t )-(k t+1 /k t). Let δ* denote the unconditional mean of the time series [1+(dkt/ kt)-(kt+1/kt)]; that is, This is the moment restriction that underlies our estimate of . The first-order necessary condition for hours worked requires that, for all t, the marginal productivity of hours times the marginal utility of consumption equals the marginal disutility of working. This implies the condition ␥ =(1-)(yt/nt)/ [ctV’(N-nt)] for all t. Let ␥ * denote the unconditional expected value of the time series on the right side of that condition; that is, (19) We identify ␥ with a consistent estimate of the parameter ␥ *. Next, consider the random variable Here ⌬ denotes the first-difference operator. Under the null hypothesis of balanced growth, =Et, the unconditional growth rate of output. Therefore, (20) (17) We identify ␦ with a consistent estimate of the parameter ␦*. The social planner’s first-order necessary condition for capital accumulation requires that the time-t expected value of the marginal The relations in (20) summarize the moment restrictions underlying our estimators of and . Our assumptions regarding the stochastic process generating government consumption imply the unconditional moment restrictions: These moment restrictions can be used to estimate , and µ. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors AGGREGATE LABOR MARKET FLUCTUATIONS 438 THE AMERICAN ECONOMIC REVIEW JUNE 187 1992 struction, H P-filtered data have a zero mean. Equations (23) consist of six unconditional moment restrictions involving the six elements of ⌿2. These restrictions can be summarized as (21) (24) Equations (17)–(21) consist of eight unconditional moment restrictions involving the eight elements of ⌿ 1 . These can be summarized as (22) where is the true value of ⌿ 1 and H1, t(⌿ 1) is the 8×1 random vector which has as its elements the left sides of (17)–(21) before expectations are taken. 2. The Unconditional Moments Underlying Our Estimator of ⌿ 2.—Our estimator of the elements of ⌿ 2 coincides with standard second-moment estimators. We find it convenient to map these into the GMM framework. The first-moment conditions we use are In (23) we have used the fact that, by con(23) where is the true value of ⌿ 2 and H2; t(⌿ 2) is the 6×1 vector-valued function which has as its elements the left sides of (23) before expectations are taken. In order to discuss our estimator, it is convenient to define the 14×1 parameter vector ⌿ =[⌿ ⌿ 1 ⌿ 2 ]′ and the 14×1 vector-valued function With this notation, the unconditional moment restrictions (22) and (24) can be written as (25) where the vector of true values of ⌿ . Let gT denote the 14×1 vector-valued function (26) ⌿ 0) is a stationary Our model implies that Ht(⌿ and ergodic stochastic process. Since gT(·) has the same dimension as ⌿ , it follows from L.Hansen (1982) that the estimator ⌿ T, defined by the condition 0, is consistent for ⌿ 0. Let D T denote the matrix of partial derivatives (27) ⌿ T. It then follows from reevaluated at ⌿ =⌿ sults in L.Hansen (1982) that a consistent estimator of the variance-covariance matrix of is given by T (28) © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 188 SOME EXTENSIONS VOL. 82 NO. 3 CHRISTIANO AND EICHENBAUM: CURRENT RBC THEORIES Here, ST is a consistent estimator of the spectral density matrix of Ht(⌿0) at frequency zero.9 B. Testing Now we describe how a Wald-type test statistic described in Eichenbaum et al. (1984) and Whitney K.Newey and Kenneth D.West (1987) can be used to assess formally the plausibility of the model’s implications for subsets of the second moments of the data. Our empirical analysis concentrates on assessing the model’s implications for the labor-market moments, [corr(y/n), n/y/n].10 Here, we describe our procedure for testing this set of moments, a procedure which can be used for any finite set of moments. Given a set of values for ⌿1, our model implies particular values for [corr(y/n), n/y/ n ] in population. We represent this relationship by the function f that maps into : (29) 439 Here, f1(⌿1) and f2(⌿2) denote the model’s implication for corr(y/n, n) and n/y/n in population, conditional on the model parameters, ⌿1. The function f(·) is highly nonlinear in ⌿1 and must be computed using numerical methods. We use the spectral technique described in Christiano and Eichenbaum (1990b). Let A be the 2×14 matrix composed of zeros and ones with the property that (30) and let (31) Under the null hypothesis that the model is correctly specified, (32) If our data sample were large, then and (32) could be tested by simply comwith a 2×1 vector of zeros. Howparing ever, need not be zero in a small sample, To test because of sampling uncertainty in (32), then, we need the distribution of under the null hypothesis. Taking a first-order Taylor-series approximation of about ⌿0 yields (33) 9 Let denote the spectral density matrix of Ht(⌿0) at frequency zero. Our estimator of S0, ST, uses the damped, truncated covariance estimator discussed by Eichenbaum and Hansen (1990). The results we report were calculated by truncating after six lags. Strictly speaking, HP-filtered data do not satisfy the Eichenbaum and Hansen (1990) assumption that S0 be nonsingular. This is because our model implies that data need to be differenced only once to induce stationarity, while the results of King and Rebelo (1988) show that the HP filter differences more than once. We think this is not a serious problem from the perspective of asymptoticdistribution theory. This is because our numerical results would have been essentially unchanged had we worked with a version of the HP filter in which the extra unit roots were replaced by roots arbitrarily close to 1. Then, the Eichenbaum and Hansen (1990) analysis would apply without caveat. What the small-sample properties are in the presence of unit roots in the data-generating process remains an open and interesting question. 10 Our formal test does not include n/y because this is an exact function of [corr(y/n, n), n/y/n]. To see this, let b=n/y/n and c=corr(y/n, n). Then, after some algebraic manipulation, n / y = b / ( 1 + 2 c b + b 2 ) 1/2 . It follows that a consistent estimator of the is given variance-covariance matrix of by (34) An implication of results in Eichenbaum et al. (1984) and Newey and West (1987) is that the test statistic (35) © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors AGGREGATE LABOR MARKET FLUCTUATIONS 440 THE AMERICAN ECONOMIC REVIEW is asymptotically distributed as a chi-square random variable with two degrees of freedom. We used this fact to test the null hypothesis (32). C. The Data Next we describe the data used in our analysis. In all of our empirical work, private conwas measured as quarterly real sumption, expenditures on nondurable consumption goods plus services plus the imputed service flow from the stock of durable goods. The first two time series came from the U.S. Department of Commerce’s Survey of Current Business (various issues). The third came from the data base documented in Flint Brayton and Eileen Mauskopf (1985). Government consumption, gt, was measured by real government purchases of goods and services minus real investment of government (federal, state, and local).11 A measure of government investment was provided to us by John C.Musgrave of the U.S. Bureau of Economic Analysis. This measure is a revised and updated version of the measure discussed in Musgrave (1980). Gross investment, dkt, was measured as privatesector fixed investment plus government fixed investment plus real expenditures on durable goods. The capital-stock series, kt, was chosen to match the investment series. Accordingly, we measured k t as the sum of the stock of consumer durables, producer structures and equipment, government and private residential capital, and government nonresidential capital. plus Gross output, yt, was measured as gt plus dkt plus time-t inventory investment. Given our consumption series, the difference between our measure of gross output and the one reported in the Survey of Current Business is 189 JUNE 1992 that ours includes the imputed service flow from the stock of consumer durables but excludes net exports. We used two different measures of hours worked and average productivity. Our first measure of hours worked corresponds to the one constructed by G.Hansen (1984) which is based on the household survey conducted by the U.S. Department of Labor. A corresponding measure of average productivity was constructed by dividing our measure of gross output by this measure of hours. For convenience, we refer to this measure of nt and (y/n)t as household hours worked and household productivity. A potential problem with our measure of household average productivity is that gross output covers more sectors than does the household hours data (for details, see appendix 1 of Christiano and Eichenbaum [1988]). In order to investigate the quantitative impact of this problem, we considered a second measure of hours worked and productivity which covers the same sectors: output per hour of all persons in the nonagricultural business sector (CITIBASE mnemonic LBOUTU) and per capita hours worked by wage and salary workers in private nonagricultural establishments as reported by the U.S. Department of Labor (Bureau of Labor Statistics, IDC mnemonic HRSPST). For convenience, we refer to this measure of nt and (y/n)t as establishment hours worked and establishment productivity. All data, except those for (y/n) t, were converted to per capita terms using an efficiency-weighted measure of the population. The data cover the period from the third quarter of 1955 through the fourth quarter of 1983 (1955:3–1983:4) (for further details on the data, see Christiano [1987b, 1988]). III. Empirical Results 11 It would be desirable to include in gt a measure of the service flow from the stock of government-owned capital, since government capital is included in our measure of kt. Unfortunately, we know of no existing measures of that service flow. This contrasts with household capital, for which there are estimates of the service flow from housing and the stock of consumer durables. The first is included in the official measure of consumption of services, and the second is reported by Brayton and Mauskopf (1985). In this section, we report our empirical results. Subsection A discusses the results obtained using the household data while Subsection B presents results based on the establishment data. In each case, our results © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 190 SOME EXTENSIONS VOL. 82 NO. 3 CHRISTIANO AND EICHENBAUM: CURRENT RBC THEORIES 441 TABLE 1—MODEL PARAMETER E STIMATES (AN D STAN DARD E RRORS) G ENERATED BY THE HOUSEHOLD DATA SET Notes: Standard errors are reported in parentheses only for estimated parameters. Other parameters were set a priori. are presented for four models. These correspond to versions of the model in Section II with V given by (3) or (4) and ␣=1 or 0. We refer to the model with V given by (3) and ␣=1 as our base model. A. Results for the Household Data Table 1 reports our estimates of ⌿1 along with standard errors for the different models. (We report the corresponding equilibrium decision rules in Christiano and Eichenbaum [1990a].) Table 2 documents the implications of our estimates of ⌿1 for various first moments of the data. To calculate these, we used the fully parameterized models to simulate 1,000 time series, each of length 113 (the number of observations in our data set). First moments were calculated on each synthetic data set. Table 2 reports the average value of these moments across synthetic data sets as well as estimates of the corresponding first moments of the data. As can be seen, all of the models do extremely well on this dimension. This is not surprising, given the nature of our estimator of ⌿ 1 . Notice that the models k t, predict the same mean growth rate for g t, and y t . This prediction reflects the balanced-growth properties of our models. This prediction does not seem implausible given the point estimates and standard errors reported in Table 2.12 The models also pre12 The large standard error associated with our estimate of the growth rate of gt may well reflect a break in the data around 1970. For example, the sample © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors AGGREGATE LABOR MARKET FLUCTUATIONS 442 THE AMERICAN ECONOMIC REVIEW 191 JUNE 1992 TABLE 2—SELECTED FIRST-MOMENT PROPERTIES, HOUSEHOLD DATA SET Notes: Numbers in the columns under the “model” heading are averages, across 1,000 simulated data sets, each with 113 observations, of the sample average of the variables in the first column. Numbers in parentheses are the standard deviations across data sets. The last column reports empirical averages, with standard errors in parentheses. dict that the unconditional growth rate of nt will be zero. This restriction also seems reasonably consistent with the data. Table 3 displays estimates of a subset of the second moments of the household data, as well as the analog model predictions. All of the models do reasonably well at matching the dk/y, g/y, and estimated values of y. Interestingly, introducing government into the analysis (i.e., moving from ␣=1 to ␣=0) actually improves the performance of the average of the growth rate of gt between 1955:2 and 1970:1 is 0.0060, whereas between 1970:1 and 1984:1 it is –0.0018. models with respect to dk/y, and g/ y but has little impact on their predictions for y. The models do not do well, however, at matching the volatility of hours worked relative to output (n/y). Not surprisingly, incorporating government into the analysis (␣=0) generates additional volatility in nt, as does allowing for indivisibilities in labor supply. Indeed, the quantitative impact of these two perturbations to the base model (divisible labor with ␣=1) is similar. Nevertheless, even when both effects are operative, the model still underpredicts the volatility of nt relative to yt. Similarly, allowing for non-convexities in labor supply and introducing government into the analysis improves the model’s performance © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 192 SOME EXTENSIONS VOL. 82 NO. 3 CHRISTIANO AND EICHENBAUM: CURRENT RBC THEORIES 443 TABLE 3—SECOND-MOMENT PROPERTIES AFTER HP-DETRENDING MODELS ESTIMATED USING THE HOUSEHOLD DATA SET Notes: All of the statistics in this table are computed after first logging and then detrending the data using the Hodrick-Prescott filter. Here, i is the standard deviation of variable i detrended in this way and corr(x, w) is the correlation between detrended x and detrended w. Numbers in the columns under the “model” heading are averages, across 1,000 simulated data sets, each with 113 observations, of the sample average of the variables in the first column. Numbers in parentheses are the standard deviations across data sets. The last column reports results for U.S. data with associated standard errors, computed as discussed in the text, in parentheses. with respect to the volatility of nt relative to yt/ nt. In fact, the model that incorporates both of these effects actually overstates the volatility of nt relative to yt/nt.13 Next we consider the ability of the different models to account for the DunlopTarshis observation. Table 3 shows that the prediction of the base model is grossly 13 These results differ in an important way from those of G.Hansen (1985). Using data processed with the HP filter, he reports that the indivisible labor model with ␣=1 implies a value of n/y/n equal to 2.7 (Hansen, 1985 table 1). This is more than twice the corresponding empirical quantity. Our version of this model (␣=1) underpredicts n/y/n by more than 20 percent. The reason for the discrepancy is that Hansen models innovations to technology as having a transient effect on zt, whereas we assume that their effect is permanent. Consequently, the intertemporal substitution effect of a shock to technology is considerably magnified in Hansen’s version of the model. inconsistent with the observed correlation between average productivity and hours worked. Introducing nonconvexities in labor supply has almost no impact on the model’s prediction for this correlation.14 However, introducing government into the analysis (␣ =0) does reduce the prediction some, at least moving it in the right direction. But not nearly enough: the models with ␣=0 still substantially overstate the correlation 14 To gain intuition into this result, consider a static version of our model, with no capital, in which the wage is equated to the marginal product of labor in each period. In that model, introducing indivisibilities can be thought of as flattening the labor-supply schedule, thereby increasing the fluctuations of hours worked relative to the wage. However, as long as the only shocks are to technology, the correlation between hours worked and the wage will still be strongly negative, regardless of the slope of labor supply. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors AGGREGATE LABOR MARKET FLUCTUATIONS 444 THE AMERICAN ECONOMIC REVIEW 193 JUNE 1992 TABLE 4—DIAGNOSTIC RESULTS WITH THE Two DATA SETS Notes: All results are based on data detrended by the Hodrick-Prescott filter. The numbers in the “U.S. data” column are point estimates based on U.S. data for the statistic. The portion of this column in panel A is taken directly from Table 3. The numbers in parentheses are the associated standard-error estimates. The numbers in the columns under the “model” heading are the probability limits of the statistics implied by the indicated model at its estimated parameter values; the numbers in parentheses are the standard errors of the discrepancy between the statistic and its associated sample value, reported in the U.S. data column. This standard error is computed by taking the square root of the appropriate diagonal element of equation (34). The numbers in brackets are the associated t statistics. The J statistic is computed using equation (35), and the number in braces is the probability that a chi-square with two degrees of freedom exceeds the reported value of the associated J statistic. between average productivity and hours worked. Panel A in Table 4 reports the results of implementing the diagnostic procedures discussed in Section II. The last row of the panel (labeled “J”) reports the statistic for testing the joint null hypothesis that the model predictions for both corr(y/n, n) and n/ y/n are true. As can be seen, this null hypothesis is overwhelmingly rejected for every version of the model. Notice also that the t statistics (given in brackets in the table) associated with corr(y/n, n) are all larger than the corresponding t statistics associated with n/y/n. This is consistent with our claim that the single most striking failure of existing RBC models lies in their implications for the Dunlop-Tarshis observation, rather than the relative volatility of hours worked and average productivity. B. Results Based on Establishment Data There are at least two reasons to believe that the negative correlation between hours worked and average productivity reported above is spurious and reflects measurement error. One potential source of distortion lies in the fact that gross output covers more sectors than household hours. The other potential source of distortion is that household hours data may suffer from classical © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 194 SOME EXTENSIONS VOL. 82 NO. 3 CHRISTIANO AND EICHENBAUM: CURRENT RBC THEORIES 445 TABLE 5—MODEL PARAMETER ESTIMATES (AND STANDARD ERRORS) G ENERATED BY THE ESTABLISHMENT DATA SET Notes: Standard errors are reported (in parentheses) only for estimated parameters. Other parameters were set a priori. measurement error. Classical measurement error in n will bias standard estimates of corr(y/ n, n) downward. In order to investigate the quantitative impact of the coverage problem, we redid our analysis using establishment hours worked and establishment average productivity. An important virtue of these measures is that they cover the same sectors. With these data, the estimated value of corr(y/n, n) becomes positive: 0.16 with a standard error of 0.08. This result is consistent with the view that the negative correlation reported in panel A of Table 4 reflects, in part, coverage problems with the household data. Interestingly, our estimate of n/y/n is also significantly affected by the move to the new data set. This increases to 1.64 with a standard error of 0.16. Thus, while the models’ performance with respect to the Dunlop-Tarshis observation ought to be enhanced by moving to the new data set, it ought to deteriorate with respect to the relative volatility of hours worked and output per hour. Therefore, the net effect of the new data set on overall inference cannot be determined a priori. To assess the net impact on the models’ performance, we reestimated the structural parameters and redid the diagnostic tests discussed in Section II. The new parameter estimates are reported in Table 5. The data used to generate these results are the same as those underlying Table 1, with two exceptions. One has to do with the calculations associated with the intratemporal Euler equation, that is, the third element of Ht(·). Here we used our new measure of average productivity, which is actually an index. This measure of average productivity was scaled so that the sample mean of the transformed index coincides with the sample mean of our measure of yt divided by establishment hours. The other difference is that, apart from the calculations involving © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors AGGREGATE LABOR MARKET FLUCTUATIONS 446 THE AMERICAN ECONOMIC REVIEW y t/n t, we measured n t using establishment hours. The new second-moment implications, with the exception of those pertaining to y, corr(y/n, n), and n/y/n, are very similar to those reported in Table 3. The new values of y are 0.013 (0.0017) and 0.014 (0.002) for the versions of the divisible-labor model without government ( ␣ =1) and with government (␣=0), respectively, and 0.015 (0.0019) and 0.016 (0.002) for the versions of the indivisiblelabor model without and with government. (Numbers in parentheses are standard deviations, across synthetic data sets.) The fact that these values are all lower than those in Table 3 primarily reflects our finding that the variance of the innovation to the Solow residual is lower with the establishment hours data. The results of our diagnostic tests are summarized in panel B of Table 4. Notice that, for every version of the model, the J statistic in panel B is lower than the corresponding entry in panel A. Nevertheless, as long as government is not included (i.e., when a=1), both versions of the model are still rejected at essentially the zero-percent significance level. However, this is no longer true when government is added (when ␣= 0). Then, we cannot reject the indivisible labor model at even the 15-percent significance level. To understand these results, we first consider the impact of the new data set on inference regarding the correlation between hours worked and average productivity. Comparing the ␣=0 models in panels A and B of Table 4, we see a dramatic drop in the t statistics (the bracketed numbers there). There are two principal reasons for this improvement. The most obvious reason is that (y/n, n) is positive in the new data set (0.16), while it is negative in the old data set (-0.20). In this sense, the data have moved toward the model. The other reason for the improved performance is that the new values of generate a smaller value for corr(y/n, n). For example, in the indivisible-labor model with 195 JUNE 1992 ␣=0, corr(y/n, n) drops from 0.737 to 0.575. In part, this reflects the new values of and Consider first. With the household data set, is 0.96 (after Founding) for all of the models; with the establishment data set, is 0.98 (after rounding). As we emphasized in Section I, increases in ρ are associated with decreases in the correlation between yt/nt and nt.15 Next consider With the establishment data, the estimates of γ are consistently larger than we obtained with the household data.16 For example, in the indivisible-labor model with was 0.00374; now As we noted in Section I, the impact of a change in γ on corr(y/n, n) cannot be determined a priori. As it turns out, the increase in contributes to a drop in these statistics.17 We now examine the impact of the establishment data on inference regarding the relative volatility of hours worked and average productivity. Comparing panels A and B of Table 4, we see that in all cases but one, the t statistics rise. In the exceptional case, that is, the indivisible-labor model with ␣=0, the change is very small. Three factors influence the change in these t statistics. First, the point estimate of n / y/n is larger with the establishment data. Other things equal, this 15 Consistent with this relationship, corr(y/n, n)=0.644 in the indivisible-labor model with ␣=0, when it is evaluated at the parameter values in Table 1 except with ρ set to 0.98. 16 To see why the establishment data set generates a higher value of it is convenient to concentrate on the divisible-labor model. The parameter is invariant to which data set or model is used. In practice, our estimator of is approximately where c/y denotes the sample average of and N/n denotes the sample average of N/nt. Obviously, is a decreasing function of n. The value of n with the household data set is 320.4, and the implied value of n/N is 0.23. With the establishment data set, n=257.7, and the implied value of n/N is 0.19. Our estimates of γ are different from the one used by Kydland and Prescott (1982). This is because Kydland and Prescott deduce a value of γ based on the assumption that n/N=0.33. In defending this assumption, Prescott (1986 p. 15) says that “[Gilbert R.] Ghez and [Gary S.] Becker (1975) find that the household allocates approximately one-third of its productive time to market activities and two-thirds to nonmarket activities.” We cannot find any statement of this sort in Ghez and Becker (1975). 17 For example, in the indivisible labor model with ␣=0 evaluated at the parameter estimates in Table 1, but with γ increased to 0.0046, corr(y/n, n)=0.684. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 196 SOME EXTENSIONS VOL. 82 NO. 3 CHRISTIANO AND EICHENBAUM: CURRENT RBC THEORIES hurts the empirical performance of all the models, except the indivisible-labor model with ␣=0. Second, these statistics are estimated less precisely with the establishment data, and this contributes to a reduction in the t statistics. Third, the new parameter estimates lead to an increase in each model’s implied value of n/y/n. For example, the value of n/y/n implied by the indivisible-labor model with ␣=0 rises to 1.437 from 1.348. In part, this reflects the new values of and When we evaluate the indivisiblelabor model with ␣ =0 at the parameter estimates in Table 1, with ρ increased to its Table 5 value of 0.98, the value of n/y/n equals 1.396. The analog experiment with γ increases the value of this statistic to 1.436. Comparing panels A and B of Table 4, we see that inference about the importance of the role of government consumption appears to hinge sensitively on which data set is used. On the one hand, the household data suggest that the role of government consumption is minimal. This is because both the divisiblelabor and indivisible-labor models are rejected, regardless of whether ␣=0 or 1. On the other hand, the establishment data suggest an important role for government consumption. While the divisible-labor model is rejected in both its variants, the indivisible-labor model cannot be rejected at conventional significance levels as long as ␣=0. In Christiano and Eichenbaum (1990a), we argue that the sensitivity of inference to which data set is used is resolved once we allow for classical measurement error in hours worked. The basic idea is to assume, as in Prescott (1986), that the measurement errors in the logarithm of household and establishment hours worked are uncorrelated over time and with each other, as well as with the logarithm of true hours worked. In Christiano and Eichenbaum (1990a), we show how to estimate the parameters of the models considered here, allowing for this kind of measurement error. In addition, we did the diagnostic tests that we have discussed in this paper. The main findings can be briefly summarized as follows. First, allowing for measurement error, the 447 indivisible-labor model cannot be rejected at conventional significance levels as long as government is incorporated into the analysis. This is true regardless of whether household or establishment hours data are used. Second, the divisible-labor model continues to be rejected for both data sets, regardless of whether government is included in the analysis. Therefore, with this model of measurement error, inference is not sensitive to which measure of hours worked is used. Regardless of whether household or establishment hours data are used, incorporating government into the analysis substantially improves the empirical performance of the indivisible-labor model. In Christiano and Eichenbaum (1990a), we also present evidence that the plausibility of the divisible-labor model with government is affected by the choice of stationarity-inducing transformation. In particular, there is substantially less evidence against that model with ␣=0 when the diagnostic tests are applied to growth rates of the establishment hours data set and measurement error is allowed for. IV. Concluding Remarks Existing RBC theories assume that the only source of impulses to postwar U.S. business cycles are exogenous shocks to technology. We have argued that this feature of RBC models generates a strong positive correlation between hours worked and average productivity. Unfortunately, this implication is grossly counterfactual, at least for the postwar United States. This led us to conclude that there must be other quantitatively important shocks driving fluctuations in aggregate U.S. output. We have focused on assessing the importance of shocks to government consumption. Our results indicate that, when aggregate demand shocks arising from stochastic movements in government consumption are incorporated into the analysis, the model’s empirical performance is substantially improved. Two important caveats about our empirical results should be emphasized. One has to do © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors AGGREGATE LABOR MARKET FLUCTUATIONS 448 THE AMERICAN ECONOMIC REVIEW with our implicit assumption that public and private capital are perfect substitutes in the aggregate production function. Some researchers, including most prominently Aschauer (1989), have argued that this assumption is empirically implausible. To the extent that these researchers are correct, and to the extent that public-investment shocks are important, our assumption makes it easier for our model to account for the DunlopTarshis observation. This is because these kinds of shocks have an impact on the model similar to technology shocks, and they contribute to a positive correlation between hours worked and productivity. The other caveat has to do with another implicit assumption: that all taxes are lump-sum. We chose this strategy in order to isolate the role of shocks to government consumption per se. We leave to future research the important task of incorporating distortionary taxation into our framework. How this would affect our model’s ability to account for the DunlopTarshis observation is not clear. Recent work by R.Anton Braun (1989) and Ellen R.McGrattan (1991) indicates that randomness in marginal tax rates enhances the model on this dimension. However, some simple dynamic optimal-taxation arguments suggest the opposite. Suppose, for example, that it is optimal for the government to increase distortionary taxes on labor immediately in response to a persistent increase in government consumption. This would obviously reduce the positive employment effect of an increase in government consumption. Still, using a version of our divisible-labor model, V.V.Chari et al. (1991) show that this last effect is very small. In their environment, introducing government into the analysis enhances the model’s overall ability to account for the Dunlop-Tarshis observation. In any event, if it were optimal for the government to increase taxes with a lag, we suspect that this type of distortionary taxation would actually enhance the model’s empirical performance. REFERENCES Aiyagari, S.Rao, Christiano, Lawrence J. and Eichenbaum, Martin, “The Output, Em- 197 JUNE 1992 ployment, and Interest Rate Effects of Government Consumption,” Discussion Paper No. 25, Institute for Empirical Macroeconomics (Federal Reserve Bank of Minneapolis and University of Minnesota), 1990. Aschauer, David Alan, “Fiscal Policy and Aggregate Demand,” American Economic Review, March 1985, 75, 117–27. _____, “Does Public Capital Crowd Out Private Capital?” Journal of Monetary Economics, September 1989, 24, 171–88. Barro, Robert J., “Output Effects of Government Purchases,” Journal of Political Economy, December 1981, 89, 1086–1121. _____, “Government Purchases and Public Services,” in Robert J.Barro, Macroeconomics, 2nd Ed., New York: Wiley, 1987, pp. 307–39. ______, and Grossman, Herschel I., “A General Disequilibrium Model of Income and Employment,” American Economic Review, March 1971, 61, 82–93. Bencivenga, Valerie R., “An Econometric Study of Hours and Output Variation with Preference Shocks,” International Economic Review, 1992 (forthcoming). Benhabib, Jess, Rogerson, Richard and Wright, Randall, “Homework in Macroeconomics: Household Production and Aggregate Fluctuations,” Journal of Political Economy, December 1991, 6, 1166–81. Blanchard, Olivier Jean and Fischer, Stanley, Lectures on Macroeconomics, Cambridge, MA: MIT Press, 1989. Braun, R.Anton, “The Dynamic Interaction of Distortionary Taxes and Aggregate Variables in Postwar U.S. Data,” unpublished manuscript, University of Virginia, 1989. Brayton, Flint and Mauskopf, Eileen, “The MPS Model of the United States Economy,” unpublished manuscript, Board of Governors of the Federal Reserve System, Division of Research and Statistics, Washington, DC, 1985. Chari, V.V., Christiano, Lawrence J. and Kehoe, Patrick J., “Optimal Fiscal Policy in a Business Cycle Model,” Research Department Working Paper No. 465, Federal Reserve Bank of Minneapolis, 1991. Christiano, Lawrence J., (1987a) “Dynamic Properties of Two Approximate Solutions to a Particular Growth Model,” Research Department Working Paper No. 338, Federal Reserve Bank of Minneapolis, 1987. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 198 SOME EXTENSIONS VOL. 82 NO. 3 CHRISTIANO AND EICHENBAUM: CURRENT RBC THEORIES ____, (1987b) Technical Appendix to “Why Does Inventory Investment Fluctuate So Much?” Research Department Working Paper No. 380, Federal Reserve Bank of Minneapolis, 1987. ____, “Why Does Inventory Investment Fluctuate So Much?” Journal of Monetary Economics, March/May 1988, 21, 247–80. ____ and Eichenbaum, Martin, “Is Theory Really Ahead of Measurement? Current Real Business Cycle Theories and Aggregate Labor Market Fluctuations,” National Bureau of Economic Research (Cambridge, MA) Working Paper No. 2700, 1988. ____ and _____, (1990a) “Current Real Business Cycle Theories and Aggregate Labor Market Fluctuations,” Discussion Paper No. 24, Institute for Empirical Macroeconomics (Federal Reserve Bank of Minneapolis and University of Minnesota), 1990. ___ and ____, (1990b) “Unit Roots in Real GNP: Do We Know, and Do We Care?” CarnegieRochester Conference Series on Public Policy, Spring 1990, 32, 7–61. Dunlop, John T., “The Movement of Real and Money Wage Rates,” Economic Journal, September 1938, 48, 413–34. Eichenbaum, Martin and Hansen, Lars Peter, “Estimating Models With Intertemporal Substitution Using Aggregate Time Series Data,” Journal of Business and Economic Statistics, January 1990, 8, 53–69. ____, ____ and Singleton, Kenneth J., Appendix to “A Time Series Analysis of Representative Agent Models of Consumption and Leisure Under Uncertainty,” unpublished manuscript, Northwestern University, 1984. Fischer, Stanley, “Long-Term Contracts, Rational Expectations, and the Optimal Money Supply Rule,” Journal of Political Economy, February 1977, 85, 191–205. Ghez, Gilbert R. and Becker, Gary S., The Allocation of Time and Goods Over the Life Cycle, New York: National Bureau of Economic Research, 1975. Greenwood, Jeremy and Hercowitz, Zvi, “The Allocation of Capital and Time Over the Business Cycle,” Journal of Political Economy, December 1991, 6, 1188–1214. Hansen, Gary D., “Fluctuations in Total Hours Worked: A Study Using Efficiency Units,” working paper, University of Minnesota, 1984. 449 ____, “Indivisible Labor and the Business Cycle,” Journal of Monetary Economics, November 1985, 16, 309–27. Hansen, Lars Peter, “Large Sample Properties of Generalized Method of Moments Estimators,” Econometrica, July 1982, 50, 1029–54. Hodrick, Robert J. and Prescott, Edward C., “Post-War U.S. Business Cycles: An Empirical Investigation,” unpublished manuscript, Carnegie Mellon University, 1980. Keynes, John Maynard, The General Theory of Employment, Interest and Money, New York: Harcourt Brace, 1935. King, Robert G. and Rebelo, Sergio T., “Low Frequency Filtering and Real Business Cycles,” unpublished manuscript, University of Rochester, 1988. Kormendi, Roger C., “Government Debt, Government Spending, and Private Sector Behavior,” American Economic Review, December 1983, 73, 994–1010. Kydland, Finn E. and Prescott, Edward C., “A Competitive Theory of Fluctuations and the Feasibility and Desirability of Stabilization Policy,” in Stanley Fischer, ed., Rational Expectations and Economic Policy, Chicago: University of Chicago Press, 1980, pp. 169–87. ____ and ____, “Time to Build and Aggregate Fluctuations,” Econometrica, November 1982, 50, 1345–70. ____ and ____, “The Workweek of Capital and Its Cyclical Implications,” Journal of Monetary Economics, March/ May 1988, 21, 343–60. Long, John B., Jr., and Plosser, Charles I., “Real Business Cycles,” Journal of Political Economy, February 1983, 91, 39–69. Lucas, Robert E., Jr., Studies in Business-Cycle Theory, Cambridge, MA: MIT Press, 1981. McCallum, Bennett, Monetary Economics: Theory and Policy, New York: Macmillan, 1989. McGrattan, Ellen R., “The Macroeconomic Effects of Distortionary Taxation,” Discussion Paper No. 37, Institute for Empirical Macroeconomics (Federal Reserve Bank of Minneapolis and University of Minnesota), 1991. Modigliani, Franco, “The Monetarist Controversy or, Should We Forsake Stabilization Policies?” American Economic Review, March 1977, 67, 1–19. Musgrave, John C., “Government-Owned Fixed Capital in the United States, 1925–79,” Survey of Current Business, March 1980, 60, 33–43. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors AGGREGATE LABOR MARKET FLUCTUATIONS 450 THE AMERICAN ECONOMIC REVIEW Newey, Whitney K. and West, Kenneth D., “A Simple, Positive Semi-definite, Heteroskedasticity and Autocorrelation Consistent Covariance Matrix,” Econometrica, May 1987, 55, 703–8. Phelps, Edmund S. and Winter, Sidney G., Jr., “Optimal Price Policy under Atomistic Competition,” in Edmund S.Phelps, ed., Microeconomic Foundations of Employment and Inflation Theory, New York: Norton, 1970, pp. 309–37. Prescott, Edward C., “Theory Ahead of Business Cycle Measurement,” Federal Reserve Bank of Minneapolis Quarterly Review, Fall 1986, 10, 9–22. 199 JUNE 1992 Rogerson, Richard, “Indivisible Labor, Lotteries and Equilibrium,” Journal of Monetary Economics, January 1988, 21, 3–16. Sargent, Thomas J., Macroeconomic Theory, 2nd Ed., New York: Academic Press, 1987. Shapiro, Matthew D. and Watson, Mark W., “Sources of Business Cycle Fluctuations,” National Bureau of Economic Research (Cambridge, MA) Working Paper No. 2589, 1988. Tarshis, Lorie, “Changes in Real and Money Wages,” Economic Journal, March 1939, 49, 150–4. Survey of Current Business, Washington, DC: U.S. Department of Commerce, various issues. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 200 CHAPTER 11 The Inflation Tax in a Real Business Cycle Model By THOMAS F.COOLEY AND GARY D.HANSEN* Money is incorporated into a real business cycle model using a cash-in-advance constraint. The model economy is used to analyze whether the business cycle is different in high inflation and low inflation economies and to analyze the impact of variability in the growth rate of money. In addition, the welfare cost of the inflation tax is measured and the steady-state properties of high and low inflation economies are compared. Current controversies in business cycle theory have much in common with the macroeconomic debates of the 1960s. Twenty years ago Milton Friedman and Walter Heller debated the issue of whether “money matters.” In the ensuing years the methods of business cycle research have changed dramatically but the questions have remained much the same. In particular, the issue of how much money matters is as timely now as it was when Friedman and Heller discussed it. In this paper we take the question of whether money matters to mean three things: does money and the form of the money supply rule affect the nature and amplitude of the business cycle? how does anticipated inflation affect the long-run values of macroeconomic variables? and, what are the welfare costs associated * W.E.Simon Graduate School of Management and Department of Economics, University of Rochester, Rochester, NY 14627 and Department of Economics, University of California, Los Angeles, CA 90024, respectively. We would like to acknowledge helpful comments from Steve LeRoy, David I.Levine, Bob Lucas, Bennett McCallum, Ellen McGrattan, Seonghwan Oh, Ed Prescott, Kevin Salyer, Tom Sargent, Bruce Smith, three anonymous referees, and participants in the Northwestern University Summer Research Conference, August 1988. Earlier versions of this paper were titled “The Inflation Tax and the Business Cycle.” The first author acknowledges the support of the John M.Olin Foundation. with alternative money supply rules? These are quite different questions and each implies a distinct sense in which money can affect the economy. Herein we describe a model economy that can be used to address these sorts of questions. The setting is similar to one suggested by Robert Lucas (1987) where money is held due to a cash-in-advance constraint. We use it to provide estimates of the welfare cost of the inflation tax and to study the effect of anticipated inflation on the characteristics of aggregate time-series. Early equilibrium business cycle models were influenced greatly by the monetarist tradition and the empirical findings of Milton Friedman and Anna Schwartz. They were models where unanticipated changes in the money supply played an important role in generating fluctuations in aggregate real variables and explaining the correlation between real and nominal variables (for example, Lucas, 1972). More recently, business cycle research has been focused on a class of models in which fluctuations associated with the business cycle are the equilibrium outcome of competitive economies that are subject to exogenous technology shocks. In these real business cycle models, as originally developed by Finn Kydland and Edward Prescott (1982) and John Long and Charles Plosser (1983), there is a complete set of contingent claims markets and money does not enter. Considering the importance attributed to money in earlier neoclassical and monetarist business cycle theories, it is perhaps surprising that these real models have been able to claim so much success in replicating the characteristics of aggregate data while abstracting from a role for money. This does not imply that money is unimportant for the evolution of real economic variables, 733 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE INFLATION TAX 734 THE AMERICAN ECONOMIC REVIEW but it is true that the exact role for money in these models is an open and somewhat controversial question. Not surprisingly, given that the correlation between money and output is a time-honored statistical regularity, the absence of money in real business cycle models has been a source of discomfort for many macroeconomists. One reaction to this, for example, by Ben Bernanke (1986) and Martin Eichenbaum and Kenneth Singleton (1986) among others, has been to reexamine the evidence that money “causes” changes in output. Another approach has been to construct models where money plays an essentially passive role but in which the money output correlation can be explained by distinguishing different roles for money (for example, inside and outside money) as in King and Plosser (1984) and Jeremy Greenwood and Gregory Huffman (1987). Yet another reaction has been to argue that there is some role for money over and above technology shocks. This argument is pursued in Lucas (1987). In this paper we study the quantitative importance of money in a real business cycle model where money is introduced in a way that emphasizes the influence on real variables of anticipated inflation operating through the inflation tax. Money can have important real effects in this setting: anticipated inflation will cause people to substitute away from activities that require cash, such as consumption, for activities that do not require cash, such as leisure. Nevertheless, this structure does not provide any role for unanticipated money or “sticky price” mechanisms, which many believe to be the most important channel of influence of money on the real economy. We analyze the consequence of the distortion due to anticipated inflation for real variables and estimate the magnitude of the welfare losses that result. In the following sections we describe, calibrate, and simulate a simple one-sector stochastic optimal growth model with a real economy identical to that studied by Gary Hansen (1985). The real time-series generated by the model fluctuate in response to exogenous technology shocks. The model inment lottery 201 SEPTEMBER 1989 that permits some agents to be unemployed. With the latter features, the model implies a degree of intertemporal substitution that is consistent with observed fluctuations without contradicting microeconomic evidence from panel studies. In addition, the indivisible labor assumption is consistent with the observation that most of the fluctuation in aggregate hours worked is due to fluctuations in employment rather than fluctuations in the average hours worked of an employed worker. Money is introduced into the model using a cash-in-advance constraint. Economies with this feature have been studied extensively by Alan Stockman (1981), Lucas (1982), Lucas and Nancy Stokey (1983, 1987) and Lars Svensson (1985). The cash-in-advance constraint applies only to the consumption good. Leisure and investment in our model are credit goods. Thus, if agents in this economy wish to reduce cash holdings in response to higher inflation, they can only do so by reducing consumption. In the next section of the paper we lay out the details of our model and describe the competitive equilibrium. Solving for an equilibrium in this economy is more difficult than in other real business cycle economies because the inefficiency imposed by the cashin-advance constraint rules out the use of invisible hand arguments based on the second welfare theorem. In Section III we describe how we solve for an equilibrium directly using a method described in Kydland (1987). In Section IV of the paper we present the results of some simulations of the model under various assumptions about the behavior of the monetary growth rate. Our purpose here is to use our model as an experimental device to study the effect of certain parameter interventions.1 We take a model whose statistical properties have been studied previously and examine how injections of money, operating through a cash-in-advance constraint, alter the conclusions derived from this purely real 1 See Thomas Cooley and Stephen LeRoy (1985) for a discussion of parameter and variable interventions. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 202 SOME EXTENSIONS VOL. 79 NO. 4 COOLEY AND HANSEN: THE INFLATION TAX economy. In this model, when money is supplied optimally, the real economy and its associated steady-state paths and cyclical characteristics are identical to those in Hansen (1985). This follows from the fact that when money is supplied optimally, the cash-in-advance constraint is not binding. By varying the rate of growth of the money supply we can study how the real allocation and the comovements among variables are altered. In addition we are able to measure the welfare costs of the inflation tax. The results of the experiments just described are easily summarized. When money is supplied according to a constant growth rate rule that implies positive nominal interest rates, individuals substitute leisure for goods, output and investment fall, and the steady-state capital stock is lower. The features of the business cycle are unchanged by these constant growth rates. We also report the results of experiments in which money is supplied not at a constant rate but erratically with characteristics that mimic historical experience. In these simulations, the cyclical behavior of real variables are altered slightly: consumption becomes more variable relative to income and the price level becomes quite volatile. In addition, the correlations between these variables and output become smaller in absolute value. It is encouraging that with these changes the cyclical properties of the model more closely match U.S. postwar experience. Using definitions described in Section IV we estimate the welfare cost due to the inflation tax of a sustained moderate (10 percent) inflation to be about 0.4 percent of GNP using M1 as the relevant definition of money and a quarter as the period over which it must be held. This is very close to estimates that have been suggested by others. We find the welfare costs to be much lower, about 0.1, when the relevant definition of money is the monetary base and the period over which it is constrained to be held is a month. Perhaps the most striking implication of our model for the steady-state behavior of economic aggregates is that employment rates should be lower in the long run in high inflation economies. This possibility, stated somewhat differently as the proposition that in the long 735 run the Phillips curve slopes upward, has been suggested by others, most notably by Friedman (1977). We present evidence that, for a cross section of developed economies during the period 1976–1985, average inflation rates and average employment rates are negatively correlated. The conclusions drawn from our simulations reflect only the costs and consequences of money that are due to the inflation tax: there are no informational problems created by the money supply process. We conclude that if money does have a major effect on the cyclical properties of the real economy it must be through channels that we have not explored here. I. A Cash-in-Advance Model with Production The economy studied is a version of the indivisible labor model of Hansen (1985) with money introduced via a cash-in-advance constraint applied to consumption. That is, consumption is a “cash good” while leisure and investment are “credit goods,” in the terminology of Lucas and Stokey (1983, 1987). In this section we describe the economy and define a competitive equilibrium. In the next section we describe how an equilibrium can be computed using a linear-quadratic approximation of the economy. We assume a continuum of identical households with preferences given by the utility function, (1) where ct is consumption and is leisure in time t. Households are assumed to be endowed with one unit of time each period and supply labor to a firm which produces the goods. Households are also engaged in accumulating capital which they rent to the firm. We assume that households enter period t with nominal money balances equal to mt-1 that are carried over from the previous period. In addition, these balances are augmented with a © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE INFLATION TAX 736 THE AMERICAN ECONOMIC REVIEW lump-sum transfer equal to (gt-1)Mt-1, where Mt is the per capita money supply in period t. The money stock follows a law of motion (2) In this paper, we study two economies. In the first, the gross growth rate of money, gt, is assumed to be constant. In the other economy, the log of the gross growth rate of the money supply evolves according to an autoregression of the form: (3) In equation (3), ξt is an iid random variable with and variance where mean is the unconditional mean of logarithm of the growth rate gt. It is assumed that gt is revealed to all agents at the beginning of period t. Households are required to use these previously acquired money balances to purchase the nonstorable consumption good. That is, a household’s consumption choice must satisfy the constraint, 203 SEPTEMBER 1989 nonconvex. However, following Richard Rogerson (1988), we convexify the economy by assuming that agents trade employment lotteries. That is, households sell contracts which specify a probability of working in a given period, πt, rather than selling their labor directly. Since all agents are identical, they will all choose the same πt. Thus, a fraction πt of the households will work h0 hours and the remaining (1-πt) households will be unemployed during period t. A lottery determines which of the households work and which do not. Thus, per capita hours worked in period t is given by (5) The market structure described above implies that the period utility function of the representative household as a function of consumption and hours worked is given by5 (4) where pt is the price level at time t. In this paper, attention is focused on examples where this constraint always holds with equality. A sufficient condition for this constraint to be binding is that the gross growth rate of money, gt, always exceeds the discount factor, β. Our examples will satisfy this condition.2 In our view this assumption is not unreasonable given the observed behavior of the actual money supply.3 As in Hansen (1985), labor is assumed to be indivisible. This means that households can work some given positive number of hours, h0<1, or not at all. They are not allowed to work an intermediate number of hours.4 Under usual market interpretations, this assumption implies that the consumption set of households is 2 It can be shown from the first-order conditions of the household’s problem that the cashin-advance constraint will be binding (the Lagrange multiplier associated with constraint (3) will be positive) if and only if Et(1/gt+1)<1/β. This condition follows from the use of log utility and the timing assumptions. 3 In addition, to relax this assumption would considerably complicate our solution procedure, forcing us to consider the possibility of both corner and interior solutions. 4 The indivisible labor assumption implies that all changes in total hours worked are due to changes in the number of workers. Although over half of the variance in total hours in the United States is unambiguously due to fluctuations in employment, there is still a significant percentage that is due to fluctuation in average hours. A model that allows for adjustment along both of these margins is studied in J.O.Cho and Cooley (1988). 5 This derivation makes use of the fact that consumption is the same whether or not the household is employed. This result, which holds in equilibrium, follows from the separability of (1) in consumption and leisure and is shown formally in Hansen (1985). It is possible to have unemployed agents consume less than employed without significantly affecting the results obtained from the model by assuming a nonseparable utility function (see Hansen, 1986). A more robust feature of this model is that utility is higher for unemployed individuals than for employed. Rogerson and Randall Wright (1988) show that this implication can be reversed if leisure is assumed to be an inferior good. It is unclear how one would reverse this implication without significantly affecting the other results obtained from the model. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 204 SOME EXTENSIONS VOL. 79 NO. 4 COOLEY AND HANSEN: THE INFLATION TAX We rewrite this as, (6) 737 an exogenous shock to technology that follows a law of motion given by (10) where In the remainder of this section, we will discuss the problem faced by a representative agent with preferences given by (6) as a stand-in for the individual household with preferences given by (1) who is subject to the labor indivisibility restriction. This representative household must choose consumption, investment (xt), and nominal money holdings subject to the following budget constraint:6 (7) where et is an iid random variable with mean 0 We assume that zt, like gt, is and variance revealed to all agents at the beginning of period t. The firm seeks to maximize profit, which is The first-order equal to conditions for the firm’s problem yield the following functions for the wage rate and rental rate of capital: (11) (12) In this equation, wt and rt are the wage rate and rental rate of capital, respectively. Investment is undertaken to augment the capital stock (kt) owned by the household. The capital stock obeys the following law of motion: (8) The firm in our economy produces output (Yt) using the constant returns to scale technology: (9) Capital letters are used to distinguish per capita variables that a competitive household takes as parametric from individual-specific variables that are chosen by the household.7 The variable zt is 6 This budget constraint incorporates the fact that consumption and investment sell at the same price even though one is a cash good and the other a credit good. This is because, from the point of view of the seller, sales of both credit goods and cash goods result in cash that will be available for spending at the same time in the following period. Although cash good sales in a given period result in cash receipts in the same period, this cash can not be spent until the next period. A change in variables is introduced so that the problem solved by the households will be stationary. That is, let In addition, let be the equilibrium maximized present value of the utility stream of the representative household who enters the period with a fraction of per capita money balances equal to and a capital stock equal to k when the aggregate state is described by z, g, and K. Implicit in the functional form of V are the equilibrium aggregate decision rules (H and X) as functions of the and the pricing function aggregate state, which is taken as given by the households. The function V must satisfy Bellman’s equation (primes denote next period values)8 (13) 7 In equilibrium these will be the same. Note that the solution to the firm’s profit maximization problem has been substituted into this problem through the functions w( ) and r( ). 8 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE INFLATION TAX 738 THE AMERICAN ECONOMIC REVIEW subject to 205 SEPTEMBER 1989 compute an equilibrium for our cash-in-advance economy.9 Kydland’s method involves computing a linear-quadratic approximation to the household’s problem (13). This dynamic programming problem is then solved by iterating on Bellman’s equation, requiring that the second equilibrium condition (refer to the above definition of equilibrium) hold at each step of this recursive procedure. In the remainder of this section, we outline in more detail how this procedure is implemented in our particular case. The first step is to substitute the nonlinear constraints, (14) and (15), into the household’s utility function (6). This is done by first eliminating c by substituting (15) into (14) and (6). The resulting budget constraint is (14) (15) (16) (17) (18) (19) and c, x, nonnegative and 0≤h≤1. In addition, X, H, and are given functions of (z, g, K). A stationary competitive equilibrium for this economy consists of a set of decision rules, c(s), x(s), and h(s) (where a set of aggregate decision rules, X(S) and H(S) and a (where S=(z, g, K)), a pricing function value function V(s) such that: (i) the functions V, X, H, and satisfy (13) and h are the associated set of decision and c, x, rules; when k=K and (ii) x=X, h=H, and and (iii) the functions c(s) and x(s) satisfy c(s)+x(s)=Y(S) for all s. (20) Because of the constant returns to scale technology, requiring that the functions w and r be of the form (11) and (12) guarantees that equilibrium condition (iii) is satisfied. The constraint (20) can be substituted into the utility function (6) by eliminating h. However, we must first eliminate H. This is done by aggregating (20) and solving for H. Using (11) and (12), this implies (21) II. Solution Method In Hansen (1985) it was possible to compute an equilibrium indirectly by solving for the (unique) equal weight Pareto optimal allocation and invoking the second welfare theorem. In order to obtain an analytic solution to the problem, a linear-quadratic approximation to this nonlinear problem was formed, making it possible to compute linear decision rules. Unfortunately, it is not possible to invoke the second welfare theorem to compute an equilibrium for the economy studied in this paper. This is because money introduces a “wedge of inefficiency” (in the words of Lucas, 1987) that forces one to solve for an equilibrium directly. To get around this, we apply the method described in Kydland (1987) to Equation (21) can be substituted into (20), and the result substituted into (6). The return func- 9 This method is similar to the method of Kydland and Prescott (1977), which is described in some detail in Thomas Sargent (1981) and Charles Whiteman (1983). In addition to Kydland’s method, a number of other approaches to solving dynamic equilibrium models with distortions have been recently proposed in the literature. Examples include papers by David Bizer and Kenneth Judd (1988), Marianne Baxter (1988), and Wilbur Coleman (1988). © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 206 SOME EXTENSIONS VOL. 79 NO. 4 COOLEY AND HANSEN: THE INFLATION TAX tion for the household’s dynamic programming problem is now given by the following expression: (22) 739 relating U to S. We start with a guess for the matrix V, call it V 0 , and consider the maximization problem on the right side of (23). Once the laws of motion, (16) through (19), have been substituted into the objective, we obtain from the first-order condition for u the linear decision rule (24) In order to obtain an analytic solution to this problem, the above nonlinear return function (22) is approximated by a quadratic function in the neighborhood of the steady state of the certainty problem. This approximation technique is described in detail in Kydland and Prescott (1982). The state vector of the resulting linearquadratic dynamic programming problem is K, k)T and the individuals′ decision s=(1, z, g, In addition, (or control) vector is also the economywide variables enter the quadratic return function. Thus, after computing the quadratic approximation of (22), Bellman’s equation for the household’s problem (13) become10 (23) By imposing the equilibrium conditions, x=X, and k=K; we can obtain, from (24), a linear expression for U in terms of S that we take as our candidate. That is, we obtain (25) To compute the value function for the next iteration, we evaluate the objective function on the right side of (23) using our initial guess V0, the function relating U to S (25) and the household’s decision rule (24).11 This provides a quadratic form, sTV1s, that is used as the value function for the next iteration. This procedure is repeated until Vj+1 is sufficiently close to Vj to claim that the iterations have converged. Once this process has converged, we obtain the following equilibrium expressions for X and ( is equal to the inverse of consumption in an equilibrium where the cash-in-advance constraint is always binding): (26) (27) subject to (16)–(19) and a linear function that describes the relationship between U and S=(1, z, g, K) T perceived by the agents in the model. To solve for an equilibrium, we iterate on this quadratic version of Bellman’s equation. This procedure must involve choosing a candidate for the perceived linear function 10 This form for Bellman’s equation incorporates both certainty equivalence and the fact that the value function will be quadratic. Examples of these decision rules for particular parameterizations of the money supply rule are 11 For the parameterizations studied in this paper it is not always possible to invert the firstorder conditions to obtain an expression like (24). However, it is always possible to obtain equation (25). Therefore, when evaluating (23), we used (25) and, in place of (24) the equilibrium expressions for the components of u( and x=X). The first-order conditions are satisfied given the way in which (25) is constructed and the fact that the coefficients on k and always turn out to equal zero in these first-order conditions. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE INFLATION TAX 740 THE AMERICAN ECONOMIC REVIEW given in the Appendix. These equations, which determine investment and consumption, along with the laws of motion (16) through (18), the expression for hours worked (21), and the technology (9), are used to simulate artificial timeseries for various parameterizations of the gt process. These experiments are discussed in the next section. III. Results We use the artificial economy just described to study the interaction between money and the real sector of the economy. We first describe the cyclical behavior of our economy under various money supply rules. We then use the model to measure the welfare costs of anticipated inflation. Finally, we look for confirmation of the implied steady-state behavior of high and low inflation economies in cross-section data on several developed countries. A. Cyclical Properties Statistics summarizing the cyclical behavior of our model economy under various money supply rules, as well as statistics summarizing the cyclical behavior of actual U.S. time-series, are presented in Table 1. We will begin by describing how these statistics are computed and then proceed to interpret our results. The first panel of Table 1 shows the (percent) standard deviations of the set of endogenous variables and their correlations with output that characterize recent U.S. quarterly data. These provide some basis for comparison with the results of our experiments although we wish to stress that ours is not a data matching exercise but an experimental simulation of a model economy. We use quarterly data from 1955,3 to 198 4,1 on real G N P, consumption, investment, capital stock, hours worked, productivity, and two measures of the price level, the CPI and GNP deflator. 12 Before 12 The series for real GNP, investment, hours worked, and the price level were taken from the Citibase database. The hours series is based on 207 SEPTEMBER 1989 computing statistics, the data (both actual and simulated) are logged and detrended using the Hodrick-Prescott filter. The use of this detrending procedure enables us to maintain comparability with prior real business cycle studies by Kydland and Prescott (1982) and Hansen (1985). In order to derive results from the artificial economies, we follow Kydland and Prescott (1982) by choosing parameter values based on growth observations and the results of studies using microeconomic data. In order to make comparisions with Hansen (1985) meaningful, we set the parameters describing preferences and technology to the same values used in that study. Those values, which were chosen under the assumption that the length of a period is one quarter, are β=0.99, θ=0.36, δ=0.025, B=2.86, and γ=0.95. The standard deviation of ε, σε, is set equal to 0.00721 so that the standard deviation of the simulated output series is close to the standard deviation of the actual output series. We experiment with different values for the parameters describing the money supply process. Given a set of parameter values, simulated time-series with 115 observations (the number of observations in the data sample) are computed using the method described in the previous section. These series are then logged and filtered and summary statistics calculated. We simulate the economy 50 times and the averages of the statistics over these simulations are reported. In addition, we report the sample standard deviations of these statistics, which are given in parentheses. The columns of the second panel of Table 1 show the percent standard deviations and correlations that result from all of the simulations information from the Current Population Survey. Productivity is output divided by hours worked. The data on the capital stock include government capital stock and private capital stock (housing) as well as producers’ durables and structures. The consumption series includes nondurables and services plus an imputed flow of services from the stock of durables. The consumption and capital stock series were provided by Larry Christiano. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 208 SOME EXTENSIONS VOL. 79 NO. 4 COOLEY AND HANSEN: THE INFLATION TAX 741 TABLE 1—STANDARD DEVIATIONS IN PERCENT AND CORRELATIONS WITH OUTPUT FOR U.S. AND ARTIFICIAL ECONOMICS a The U.S. time-series reported on are real GNP, consumption of nondurables and services, plus the flow of services from durables, gross private domestic investment (all in 1982 dollars). The capital stock series includes nonresidential equipment and structures, residential structures, and government capital. The hours series is total hours for persons at work in nonagricultural industries as derived from the Current Population Survey. Productivity is output divided by hours. All series are seasonally adjusted, logged, and detrended. The output, investment, hours, and price-level series were taken from the Citibase database. The consumption and capital stock series were provided by Larry Christiano. b The percent standard deviations and correlations with output are sample means of statistics computed for each of 50 simulations. Each simulation is 115 periods long, which is the same number of periods as the U.S. sample. The sample standard deviations of these statistics are in parentheses. Each simulated time-series was logged and detrended using the same procedure applied to the U.S. sample before the statistics were calculated. of our model economy where the money supply grows at a constant rate. These results confirm that when money is supplied at a constant growth rate, even one that implies a high average inflation rate, the features of the business cycle are unaffected. In particular, the statistics summarizing the behavior of the real variables are the same as would be obtained in the same model without money—the “indivisible labor” model of Hansen (1985). The remaining two panels of Table 1 show the results of simulations with an erratic money supply. That is, we assume a money supply rule of the form (3). We calibrate this money supply process (that is, choose values for α and σζ) so that the money supply varies in a way that is broadly consistent with postwar experience. We proceed by assuming that the Fed draws money growth rates from an urn with the draws being serially correlated, as in equation (3). We determined the characteristics of that urn from data on M1 and the regression (standard errors in parentheses) © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE INFLATION TAX 742 THE AMERICAN ECONOMIC REVIEW where M1 is the average quarterly value. We intentionally averaged to smooth the data somewhat and increase the implied persistence.13 The results of this regression lead us to set a equal to 0.48 and σξ equal to 0.009. To ensure that the gross rate of money growth always exceeds the discount factor, as is required for the cash-inadvance constraint to be always binding, we draw ζt from a lognormal distribution. This implies that log(gt) will never become negative. The statistics reported in Table 1 show that volatility of the money supply has a small but significant impact on the cyclical characteristics of the economy. Virtually all the effect of volatility in the money supply is in the standard deviations of consumption and prices and their correlation with output. In particular, consumption and prices become more volatile and their correlation with output becomes smaller in absolute value. It is worth noting that the numbers in these panels are more in keeping with historical experience (see first panel) than are the results from constant growth rate economies. In addition, comparing the third and fourth panels we find that, although the price level does become more volatile, increases in the average growth rate of money has little effect on the cyclical properties of the real variables. B. Welfare Costs of the Inflation Tax In this section estimates of the welfare costs of the inflation tax are presented that are derived by comparing steady states of our growth model assuming different growth rates of the money supply.14 Measuring the welfare costs of antici13 This equation is open to criticism as a description of the historical sample. Although we cannot reject its adequacy, there may be a leftover moving average piece in the residuals. This in turn could imply that some portion of the innovation in the money growth rate is permanent. See, for example, G.William Schwert (1987). We chose to ignore this because the estimated autoregression seems to capture the features that are appropriate for our experiment. 14 A somewhat similar approach to that taken here appears in a recent paper by Jean Pierre Danthine, John Donaldson, and Lance Smith 209 SEPTEMBER 1989 pated inflation is an old issue in macroeconomics. Martin Bailey (1956) provided a classic answer to this question by considering the area under the demand curve for money, the welfare triangle, evaluated at an interest rate embodying the steady-state rate of inflation as a measure of the net loss to individuals from the inflation tax. Stanley Fischer (1981) and Robert Lucas (1981) updated Bailey’s estimates and they supply a thoughtful discussion of some of the awkward assumptions underlying the welfare triangle approach (for example, that government expenditures are financed by non-distorting taxes). They also discuss some of the subsidiary costs of inflation that are ignored by those calculations. We chose to measure the welfare costs by comparing steady states because, as explained above, the cyclical characteristics of this economy are unaffected by the average growth rate of the money stock. Thus, our discussion of welfare is based on the steady-state properties of a version of our economy where the money supply grows at a constant rate and the technology shock in equation (9) is replaced by its unconditional mean. The welfare costs for various annual inflation rates, along with the associated steady-state values for output, consumption, investment, the capital stock, and hours worked, are presented in Table 2. We show results based on two different assumptions on the length of time that the cashin-advance constraint is binding. The numbers displayed in the top panel reflect the assumption that the relevant period over which individuals are constrained to hold money is a quarter. This is consistent with the calibration of the model in the previous section. In addition, if we assume a unitary velocity as is implied by our model and if we assume that the “cash good” corresponds to consumption of nondurables and services then (1987). Their model differs from ours in that money appears directly in the utility function and they do not include labor in their model. In addition, they assume that capital depreciates fully each period. They also demonstrate a decline in welfare with inflation, but do so using simulations of their economy rather than comparing steady states. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 210 SOME EXTENSIONS VOL. 79 NO. 4 COOLEY AND HANSEN: THE INFLATION TAX 743 TABLE 2—STEADY STATES AND WELFARE COSTS ASSOCIATED WITH VARIOUS ANNUAL GROWTH RATES OF MONEY this would be consistent with defining money as M1, based on evidence from the 1980s.15 The results given in the bottom panel of Table 2 are based on the assumption that the relevant period over which individuals are constrained to hold money is a month. It turns out that monthly consumption of nondurables and services corresponds roughly to the monetary base during the 1980s. The steady states in this second panel were computed using different parameter values for the discount factor and depreciation rate of capital in order to maintain comparability to the quarterly results. The values assigned were β=0.997 and δ=0.008, which are the monthly rates that correspond to the quarterly rates assumed above. We also scale the production function to reflect monthly output levels by multiplying the right-hand side 15 This conclusion is based on the fact that the ratio of the stock of M1 to quarterly consumption of nondurables and services has been close to one since the late 1970s. Unfortunately, this result does not hold over a long period of time-the ratio has been as high as 3 early in the postwar period. The same caveat applies to the observation concerning the monetary base made below. of equation (9) by 1/3. The values for the gross growth rate of the money supply (g) that correspond to the desired annual inflation rates are also different for the monthly model. We indicate these values in the table. The welfare measure we use is based on the increase in consumption that an individual would require to be as well off as under the Pareto optimal allocation. The Pareto optimal allocation for our economy is equivalent to the equilibrium allocation for the same economy without the cash-in-advance constraint, or, equivalently, for a version of the model where the money supply grows at a rate such that the cash-in-advance constraint is never binding. It turns out that for the model studied in this paper, the cash-in-advance constraint is not binding if the gross growth rate of money is equal to the discount factor, β.16 To obtain a measure of the 16 We restrict the growth rate, g, to be greater than or equal to β. This ensures that nominal interest rates will not be negative (see Lucas and Stokey, 1987). When we set g=β, the initial price level is no longer uniquely determined. However, the real allocation and rate of inflation are uniquely determined and the allocation is Pareto optimal. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE INFLATION TAX 744 THE AMERICAN ECONOMIC REVIEW welfare loss associated with growth rates that are larger than β , we solve for ΔC in the equation (28) where is the level of utility attained (in the steady state) under the Pareto optimal allocation (g=β), and C* and H* are the steady-state consumption and hours associated with the growth rate in question (some g>β). The results of the welfare calculations expressed as a percent of steady-state real output (ΔC/Y) and steady-state real consumption (ΔC/C) are shown in the bottom rows of both panels of Table 2. The welfare cost of a moderate (10 percent) inflation is 0.387 percent of GNP when the period over which individuals are constrained is a quarter. This magnitude may be compared to the estimates of 0.3 percent provided by Stanley Fischer or 0.45 percent obtained by Robert Lucas based on an approximation of the area under a money demand function. 17 It is interesting that their exercise, which holds output constant but allows velocity to vary, yields the same answer as our exercise which holds velocity constant but allows output to vary. While an estimate of roughly 0.4 percent of GNP sounds small, at current levels of GNP it would amount to $15.2 billion of real GNP. The welfare costs of very high inflation rates, which are not uncommon throughout the world, seem extremely high. If the relevant period over which individuals are constrained is a month then the welfare costs are considerably reduced being only 0.11 percent at a 10 percent annual inflation rate and slightly more than 1.5 percent at a 400 percent annual inflation rate. Evidently the period over which individuals are constrained, and by implication the 17 Fischer and Lucas use different definitions of money (high-powered money and M1, respectively) and different estimates of the interest elasticity. 211 SEPTEMBER 1989 definition of the money balances on which individuals are taxed, make a big difference in the welfare costs of inflation. Since there is a big difference in the estimates it is worth considering what some of the biases might be. Our larger estimates come from assuming that individuals are constrained for one quarter, which is roughly consistent with assuming that the appropriate monetary aggregate is M1. However, a large part of M1 consists of checkable deposits. To the extent that these earn competitive interest they will be shielded from the inflation tax. At the other extreme, the monetary base consists of currency and reserves. Since these are clearly subject to the inflation tax, the monthly data provides a lower bound on the magnitude of the welfare loss. It seems reasonable that in economies with sustained high inflations many individuals will be able to shield themselves against the inflation tax. If the institutions did not exist to facilitate this, one would expect them to evolve in very high inflation economies. For this reason, our model may not be very reliable for analyzing hyperinflation. On the other hand these estimates abstract from many of the subsidiary costs of inflation that are believed to be important. Among these are distortions caused by nonneutralities in the tax system and adjustment costs or confusion caused by the variability of inflation. C. Steady-State Implications of Inflation As shown in Table 2, anticipated inflation has a significant influence on the steady-state path of the economy. Steady-state consumption, output, hours, investment, and the capital stock are all lower whenever the growth rate of the money supply exceeds the optimal level (g=β). The consumption of leisure increased because agents substitute this “credit good” for the consumption good in the face of a positive inflation tax on the latter. Lower hours worked leads to lower output and therefore lower consumption, investment, and capital stock. The share of output allocated to investment does not change with higher inflation. This result is obtained despite the fact that consumption is a cash good and investment is a credit good since, in the steady state, investment will provide consumption in © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 212 SOME EXTENSIONS VOL. 79 NO. 4 COOLEY AND HANSEN: THE INFLATION TAX 745 of employment rates on inflation rates. There is a statistically significant negative correlation between inflation rates and employment rates. The coefficient of the inflation rate in a regression of the employment rate on the inflation rate and a constant is -0.5 with a standard error of 0.17. The most extreme observation in the graph pertains to Chile. When that is eliminated the conclusions are essentially unchanged; the coefficient of inflation is -0.44 with a standard error 0.22. These results suggest that the phenomenon displayed in our model economy may not be counterfactual. IV. Conclusions FIGURE 1. AVERAGE EMPLOYMENT AND INFLATION RATES, 1976–1985 the future that will be subject to exactly the same inflation tax as consumption today. A striking implication of higher inflation rates in our model economy is that they are accompanied by lower employment rates.18 The “menu of choices” available to the monetary authority involves giving up low inflation only to obtain higher unemployment. This result, that the operational Phillips curve is upward sloping, is also obtained by Greenwood and Huffman (1987) for their model economy. Friedman (1977) in his Nobel lecture presented some evidence for this phenomenon by plotting data from several countries. Here we present some statistical evidence that supports the negative correlation between employment rates and inflation rates using a cross section of countries. Figure 1 shows the relation between the average rate of employment and the average rate of inflation from 1976 to 1985 for 23 countries.19 The solid line depicts the regression 18 The variable HOURS in Table 2, which corresponds to per capita hours worked, is actually the employment rate multiplied by a constant (h0), given the assumption of indivisible labor. 19 The countries are Austria, Belgium, Denmark, Finland, France, W.Germany, Greece, Ire- In this paper we incorporate an interesting paradigm for money holding, the cash-in-advance model, in a stochastic optimal growth model with an endogenous labor leisure decision. We have shown that the solution and simulation of such a model is quite tractable. The model and solution procedure provide a basis for studying the influence of inflation on the path of the real economy and its cyclical characteristics. In addition, the solution procedure we have used could be employed to study the effects of other distortions as well. We have used this model as the basis for estimating the welfare cost of the inflation tax and studying the long-run features of economies with different inflation rates. The fact that our estimates are well within the range of estimates obtained by other methods and that the empirical implications are confirmed in crosssectional data is very encouraging. This suggests to us that the approximations and simplifications we have made in writing down a tractable model of a competitive economy incorporating money may not be too serious. This is not to argue that econometric estimation of many of the parameters we have simply specified might not land, Italy, Netherlands, Norway, Portugal, Spain, Sweden, Switzerland, U K, Canada, United States, Australia, New Zealand, Japan, Chile, and Venezuela. Population data are taken from Summers and Allan Heston (1988) and the remainder of the data are taken from the International Labor Office (1987). © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE INFLATION TAX 746 THE AMERICAN ECONOMIC REVIEW yield further insights into these problems. What we find appealing about this approach is that all the features of the economy, from the decision rules to the specification of technology and preferences are explicit. Nothing is hidden. This makes it a valuable environment for experimental exercises like those considered here, and for positive exercises, for example where one would model the behavior of the monetary authority. Although we have shown that anticipated inflation can have significant effects on the longrun values of real variables, our model economy predicts that the business cycle will be the same in a high inflation economy as in a low inflation economy. When money is supplied erratically, the characteristics of the business cycle are altered somewhat. These changes in the characteristics of the cycle occur solely because of changes in allocations that result from the changing conditional expectation of inflation. Unexpected inflation has no role in this model. However, we speculate that the most important influence of money on short-run fluctuations are likely to stem from the influence of the money supply process on expectations of relative prices, as in the natural rate literature. That is, if money does have a significant effect on the characteristics of the cycle it is likely to come about because the behavior of the monetary authority has serious informational consequences for private agents. APPENDIX Decision Rules for Selected Cases Constant Growth Rate 213 SEPTEMBER 1989 Autoregressive Growth Rate REFERENCES Bailey, Martin J., “The Welfare Cost of Inflationary Finance,” Journal of Political Economy, April 1956, 64, 93–110. Baxter, Marianne, “Approximating Suboptimal Dynamic Equilibria: A Euler Equation Approach,” reproduced, University of Rochester, 1988. Bernanke, Ben S., “Alternative Explanations of the Money-Income Correlation,” CarnegieRochester Conference on Public Policy, K.Brunner and A.Meltzer, eds., Autumn 1986, 25, 49–99. Bizer, David and Judd, Kenneth, “Capital Accumulation, Risk, and Uncertain Taxation,” reproduced, University of Chicago, 1988. Cho, Jang-Ok and Cooley, Thomas F., “Employment and Hours Over the Business Cycle,” reproduced, University of Rochester, 1988. Citibank, “Citibase Economic Database,” New York: Citibank, 1978. Coleman, Wilbur, “Money, Interest, and Capital in a Cash-in-Advance Economy,” © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 214 SOME EXTENSIONS VOL. 79 NO. 4 COOLEY AND HANSEN: THE INFLATION TAX reproduced, Board of Governors of the Federal Reserve System, 1988. Cooley, Thomas F. and LeRoy, Stephen F., “Atheoretical Macroeconomics: A Critique,” Journal of Monetary Economics, November 1985, 16, 283–308. Danthine, Jean Pierce, Donaldson, John and Smith, Lance, “On the Superneutrality of Money in a Stochastic Dynamic Macroeconomic Model,” Journal of Monetary Economics, July 1987, 20, 475–500. Eichenbaum, Martin and Singleton, Kenneth, “Do Equilibrium Real Business Cycle Theories Explain Post-War U.S. Business Cycles?” in Macroeconomics Annual, Stanley Fischer, ed., Vol. I, 1986, Cambridge: MIT Press. Fischer, Stanley, “Towards an Understanding of The Costs of Inflation,” Carnegie-Rochester Conference on Public Policy, K. Brunner and A.Meltzer, eds. Autumn 1981, 15, 5– 42. Friedman, Milton, “Nobel Lecture: Inflation and Unemployment,” Journal of Political Economy, June 1977, 85, 451–72. ____, ____, and Heller, Walter W., Monetary vs. Fiscal Policy, New York: Norton, 1969. ____,____, and Schwartz, Anna J., A Monetary History of the United States, 1967–1960, Princeton: Princeton University Press, 1963. Greenwood, Jeremy and Huffman, Gregory, “A Dynamic Equilibrium Model of Inflation and Unemployment,” Journal of Monetary Economics, March 1987, 19, 203–28. Hansen, Gary D., “Indivisible Labor and the Business Cycle,” Journal of Monetary Economics, November 1985, 16, 309–28. ____, “Growth and Fluctuations,” reproduced, UCLA, 1986. International Labour Office, Yearbook of Labour Statistics, 1987, Geneva. King, Robert and Plosser, Charles, “Money, Credit, and Prices in a Real Business Cycle Economy,” American Economic Review, June 1984, 74, 363–80. Kydland, Finn E., “The Role of Money in a Competitive Model of Business Fluctuations,” reproduced, Carnegie-Mellon University, 1987. 747 ____and Prescott, Edward C., “Rules Rather Than Discretion: The Inconsistency of Optimal Plans,” Journal of Political Economy, June 1977, 85, 473–91. ___ and ____, “Time to Build and Aggregate Fluctuations,” Econometrica, November 1982, 50, 1345–70. Long, John and Plosser, Charles, “Real Business Cycles,” Journal of Political Economy, February 1983, 91, 39–69. Lucas, Robert E., Jr., “Expectations and the Neutrality of Money,” Journal of Economic Theory, April 1972, 4, 103–24. ____, “Discussion of Towards an Understanding of the Costs of Inflation: II.” CarnegieRochester Conference on Public Policy, K.Brunner and A.Meltzer, eds., Autumn 1981, 15, 43– 52. ____, “Interest Rates and Currency Prices in a Two-Country World,” Journal of Monetary Economics, November 1982, 12, 335–60. _____, Models of Business Cycles, New York: Basil Blackwell, 1987. ____ and Stokey, Nancy L., “Optimal Fiscal and Monetary Policy in an Economy Without Capital,” Journal of Monetary Economics, July 1983, 12, 55–93. ____, “Money and Interest in a Cash-in-Advance Economy,” Econometrica, May 1987, 55, 491–514. Prescott, Edward C., “Can The Cycle Be Reconciled with a Consistent Theory of Expectations?,” reproduced, Federal Reserve Bank of Minneapolis, 1983. ____, “Theory Ahead of Business Cycle Measurement,” Federal Reserve Bank of Minneapolis Quarterly Review, Fall 1986, 10, 9–22. Rogerson, Richard, “Indivisible Labor, Lotteries and Equilibrium,” Journal of Monetary Economics, January 1988, 21, 3–16. ____ and Wright, Randall, “Involuntary Unemployment in Economies with Efficient Risk Sharing,” Journal of Monetary Economics, November 1988, 22, 501–15. Sargent, Thomas, “Lecture Notes on Filtering, Control, and Rational Expectations,” reproduced, University of Minnesota, 1981. Schwert, G.William, “The Effects of Model Specification on Some Tests for Unit © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE INFLATION TAX 748 THE AMERICAN ECONOMIC REVIEW Roots,” Journal of Monetary Economics, July 1987, 20, 73–103. Stockman, Alan C., “Anticipated Inflation and the Capital Stock in a Cash-in-Advance Economy,” Journal of Monetary Economics, November 1981, 8, 387–93. Summers, Robert and Heston, Alan, “A New Set of International Comparisons of Real Product and Prices for 130 Countries, 215 SEPTEMBER 1989 1950–1985,” Review of Income and Wealth, March 1988, 34, Supplement, 1–25. Svensson, Lars E.O., “Money and Asset Prices in a Cash-in-Advance Economy,” Journal of Political Economy, October 1985, 93, 919– 44. Whiteman, Charles, Linear Rational Expectations Models: A User’s Guide, Minneapolis: University of Minnesota Press, 1983. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors Part IV The methodology of equilibrium business cycle models © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors CHAPTER 12 219 Scand. J. of Economics 93(2), 161–178, 1991 The Econometrics of the General Equilibrium Approach to Business Cycles* Finn E.Kydland Carnegie-Mellon University, Pittsburgh PA, USA Edward C.Prescott Federal Reserve Bank and University of Minnesota, Minneapolis MN, USA Abstract The founding fathers of the Econometric Society defined econometrics to be quantitative economic theory. A vision of theirs was the use of econometrics to provide quantitative answers to business cycle questions. The realization of this dream required a number of advances in pure theory—in particular, the development of modern general equilibrium theory. The econometric problem is how to use these tools along with measurement to answer business cycle questions. In this essay, we review this econometric development and contrast it with the econometric approach that preceded it. I. Introduction Early in this century American institutionists and members of the German historical school attacked—and rightfully so—neoclassical economic theory for not being quantitative. This deficiency bothered Ragnar Frisch and motivated him, along with Irving Fisher, Joseph Schumpeter, and others, to organize the Econometric Society in 1930. The aim of the society was to foster the development of quantitative economic theory—that is, the development of what Frisch labeled econometrics. Soon after its inception, the society started the journal Econometrica. Frisch was the journal’s first editor and served in this capacity for 25 years. In his editorial statement introducing the first issue of Econometrica (1933), Frisch makes it clear that his motivation for starting the Econometric * We acknowledge useful comments of Javier Diaz-Giménez on an early draft. This research was partly supported by a National Science Foundation Grant. The views expressed herein are those of the authors and not necessarily those of the Federal Reserve Bank of Minneapolis or the Federal Reserve System. Scand. J. of Economics 1991 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 220 METHODOLOGY 162 F.E.Kydland and E.C.Prescott Society was the “unification of theoretical and factual studies in economics” (p. 1). This unification of statistics, economic theory, and mathematics, he argues, is what is powerful. Frisch points to the bewildering mass of statistical data becoming available at that time, and asserts that in order not to get lost “we need the guidance and help of a powerful theoretical framework. Without this no significant interpretation and coordination of our observations will be possible” (ibid., p. 2). Frisch speaks eloquently about the interaction between theory and observation when he says “theory, in formulating its abstract quantitative notions, must be inspired to a larger extent by the technique of observation. And fresh statistical and other factual studies must be the healthy element of disturbance that constantly threatens and disquiets the theorist and prevents him from coming to rest on some inherited, obsolete set of assumptions” (ibid.). Frisch goes on to say that this mutual penetration of quantitative economic theory and statistical observation is the essence of econometrics. (ibid., p. 2). To summarize the Frisch view, then, econometrics is quantitative neoclassical theory with a basis in facts. Forty years after founding the Econometric Society, Frisch (1970) reviewed the state of econometrics. In this review he discusses what he considers to be “econometric analysis of the genuine kind” (p. 163), and gives four examples of such analysis. None of these examples involve the estimation and statistical testing of some model. None involve an attempt to discover some true relationship. All use a model, which is an abstraction of a complex reality, to address some clear-cut question or issue. It is interesting to note that, in his 1933 editorial statement, Frisch announced that each year Econometrica would publish four surveys of “the significant developments within the main fields that are of interest to the econometrician” (ibid., p. 3). These fields are general economic theory (including pure economics), business cycle theory, statistical technique, and, finally, statistical information. We find it surprising that business cycle theory was included in this list of main fields of interest to econometricians. Business cycles were apparently phenomena of great interest to the founders of the Econometric Society. Frisch’s (1933) famous, pioneering work, which appears in the Cassel volume, applies the econometric approach he favors to the study of business cycles. In this paper, he makes a clear distinction between sources of shocks on the one hand, and the propagation of shocks on the other. The main propagation mechanism he proposes is capital-starting and carry-on activities in capital construction, both of them features of the production technology. Frisch considers the implications for duration and amplitude of the cycles in a model that he calibrates using available micro data to Scand. J. of Economics 1991 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE GENERAL EQUILIBRIUM APPROACH The econometrics of the general equilibrium approach 221 163 select the numerical values for the parameters. Making the production technology with capital accumulation a central element of the theory has its parallel in modern growth theory. There are many other papers dating from the 1930s that study business cycle models. In these papers, however, and in those of the 1940s and 1950s, little progress was made beyond what Frisch had already done. The main reason was that essential theoretical tools, in particular ArrowDebreu general equilibrium theory, statistical decision theory, modern capital theory, and recursive methods had yet to be developed. The modern electronic computers needed to compute the equilibrium processes of dynamic stochastic model economies were also unavailable. Only after these developments took place could Frisch’s vision be carried out. In this paper, we review the development of econometric business cycle theory, with particular emphasis on the general equilibrium approach (which was developed later). Crucial to this development was the systematic reporting of national income and product accounts, along with time series of aggregate inputs and outputs of the business sector. Section II is a review of this important development in factual studies. In Section III we review what we call the system-of-equations approach to business cycle theory. With this approach, a theory of the business cycle is a system of dynamic equations which have been measured using the tools of statistics. Section IV is a review of the general equilibrium approach to business cycle theory. General equilibrium models have people or agents who have preferences and technologies, and who use some allocation mechanism. The crucial difference between the general equilibrium and the systemof-equations approaches is that which is assumed invariant and about which we organize our empirical knowledge. With the system-of-equations approach, it is behavioral equations which are invariant and are measured. With the general equilibrium approach, on the other hand, it is the willingness and ability of people to substitute that is measured. In Section V we illustrate the application of this econometric approach to addressing specific quantitative questions in the study of business cycles. Section VI contains some concluding comments. II. National Income and Product Accounts An important development in economics is the Kuznets-Lindahl-Stone national income and product accounts. Together with measures of aggregate inputs to the business sector, these accounts are the aggregate time series that virtually define the field of macroeconomics—which we see as concerned with both growth and business cycle fluctuations. The Kuznets-Lindahl-Stone accounting system is well-matched to the general Scand. J. of Economics © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 1991 222 METHODOLOGY 164 F.E.Kydland and E.C.Prescott equilibrium framework because there are both household and business sectors, with measures of factor inputs to the business sector and of goods produced by the business sector, as well as measures of factor incomes and expenditures on products. An examination of these time series reveals some interesting regularities— in particular, a number of ratios which remain more or less constant. These growth facts led Robert Solow to develop a neoclassical growth model which simply and elegantly rationalized these facts. Solow’s structure was not fully neoclassical, however, because the consumptionsavings decision was behaviorally determined rather than being the result of maximizing behavior subject to constraints. With the consumptionsavings decision endogenized, Solow’s growth model does become fully neoclassical, with agents’ maximizing subject to constraints and market clearing. This structure can be used to generate time series of national income and product accounts. Aggregate data present other features that are of interest to economists, such as the more volatile movements in the time series. During the 1950s and 1960s, neoclassical theory had not evolved enough to allow economists to construct computable general equilibrium models with fluctuations. Lacking the necessary tools, economists adopted an empirical approach and searched for laws of motion governing these variables. They hoped this research procedure would result in empirically determined laws which would subsequently be rationalized within the neoclassical paradigm. In the natural sciences, for example, empirically determined laws have often subsequently been rationalized at a deeper theoretical level, and it was hoped that this would also be the case in macroeconomics. In the following section we briefly review the econometrics of this approach to business cycle fluctuations. III. The System-of-Equations Approach Tjalling Koopmans, who was influenced by Frisch and might even be considered one of his students, argued forcefully in the late 1940s for what he called the econometric approach to business cycle fluctuations. At the time, it was the only econometric approach. The general equilibrium approach to the study of business cycles had yet to be developed. But since the approach Koopmans advocated is no longer the only one, another name is needed for it. As it is the equations which are invariant and measured, we label this approach the system-of-equations approach.1 1 Koopmans subsequently became disillusioned with the system-of-equations approach. When asked in the late 1970s by graduate students at the University of Minnesota in what direction macroeconomics should go, Koopmans is reported by Zvi Eckstein to have said they should use the growth model. Scand. J. of Economics 1991 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE GENERAL EQUILIBRIUM APPROACH The econometrics of the general equilibrium approach 223 165 In the 1930s, there were a number of business cycle models or theories. These logically complete theories were a dynamic set of difference equations that could be used to generate time series of the aggregate variables of interest. Notable examples include Frisch’s (1933) model in Cassel’s volume, Tinbergen’s (1935) suggestions on quantitative business cycles, and Samuelson’s (1939) multiplier-accelerator model. One problem with this class of models is that the quantitative behavior of the model depended upon the values of the coefficients of the variables included in the equations. As Haberler (1949) points out in his comment on Koopmans’ (1949) paper, the stock of cyclical models (theories) is embarrassingly large. Give any sophomore “a couple of lags and initial conditions and he will construct systems which display regular, damped or explosive oscillation…as desired” (p. 85). Pure theory was not providing sufficient discipline, and so it is not surprising that Koopmans advocated the use of the statistics discipline to develop a theory of business fluctuations. System-of-Equations Models As Koopmans (1949, p. 64) points out, the main features of the system-ofequations models are the following: First, they serve as theoretical exercises and experiments. Second, the variables involved are broad aggregates, such as total consumption, the capital stock, the price level, etc. Third, the models are “logically complete, i.e., they consist of a number of equations equal to the number of variables whose course over time is to be explained”. Fourth, the models are dynamic, with equations determining the current values of variables depending not only on current values of other variables but also on the values of beginning-of-period capital stocks and on lagged variables. Fifth, the models contain, at most, four kinds of equations, which Koopmans calls structural equations. The first type of equations are identities. They are valid by virtue of the definition of the variables involved. The second type of equations are institutional rules, such as tax schedules. The third type are binding technology constraints, that is, production functions. The final type are what Koopmans calls behavioral equations, which represent the response of groups of individuals or firms to a common economic environment. Examples are a consumption function, an investment equation, a wage equation, a money demand function, etc. A model within this framework is a system-of-equations. Another requirement, in addition to the one that the number of variables equal the number of equations, is that the system have a unique solution. A final requirement is that all the identities implied by the accounting system for the variables in the model hold for the solution to the equation system; that is, the solution must imply a consistent set of national income and product accounts. Scand. J. of Economics 1991 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 224 METHODOLOGY 166 F.E.Kydland and E.C.Prescott Statistical Measurement of Equations The behavior of these models depends crucially on the numerical magnitudes of the coefficients of the variables and of the time lags. This leads to attempts to estimate these parameters using time series of the variables being modeled. Given that the estimation of these coefficients is a statistical exercise, a probability model is an additional completeness requirement. For that purpose, a residual random disturbance vector typically is added, with one component for each behavioral equation. For statistical completeness, the probability distribution of this disturbance vector must be specified up to some set of parameters. Only then can statistical methods be applied to estimating the coefficients of the behavioral equations and the parameters of the disturbance distribution. The crucial point is that the equations of the macroeconometric model are the organizing principle of the system-ofequations approach. What is measured is the value of the coefficients of the equations. The criterion guiding the selection of the values of the coefficients is essentially the ability of the resulting system of equations to mimic the historical time series. The issue of which set of equations to estimate is settled in a similar fashion. The criterion guiding the selection of equations is in large part how well a particular set can mimic the historical data. Indeed, in the 1960s a student of business cycle fluctuations was successful if his particular behavioral equation improved the fit of, and therefore replaced, a currently established equation. The Rise and the Fall of the System-of-Equations Approach With the emergence of a consensus on the structure of the system of equations that best described the behavior of the aggregate economy, the approach advocated by Koopmans became totally dominant in the 1960s. This is well-illustrated by the following statement of Solow’s, quoted by Brunner (1989, p. 197): I think that most economists feel that the short run macroeconomic theory is pretty well in hand… The basic outlines of the dominant theory have not changed in years. All that is left is the trivial job of filling in the empty boxes [the parameters to be estimated] and that will not take more than 50 years of concentrated effort at a maximum. The reign of this system-of-equations macroeconomic approach was not long. One reason for its demise was the spectacular predictive failure of the approach. As Lucas and Sargent (1978) point out, in 1969 these models predicted high unemployment would be associated with low inflation. Counter to this prediction, the 1970s saw a combination of both high unemployment and high inflation. Another reason for the demise of this Scand. J. of Economics 1991 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE GENERAL EQUILIBRIUM APPROACH The econometrics of the general equilibrium approach 225 167 approach was the general recognition that policy-invariant behavioral equations are inconsistent with the maximization postulate in dynamic settings. The principal reason for the abandonment of the system-ofequations approach, however, was advances in neoclassical theory that permitted the application of the paradigm in dynamic stochastic settings. Once the neoclassical tools needed for modeling business cycle fluctuations existed, their application to this problem and their ultimate domination over any other method was inevitable. IV. The General Equilibrium Approach A powerful theoretical framework was developed in the 1950s and 1960s that built upon advances in general equilibrium theory, statistical decision theory, capital theory, and recursive methods. Statistical decision theory provided a logically consistent framework for maximization in a dynamic stochastic environment. This is what was needed to extend neoclassical theory, with its maximization assumption, to such environments. Another crucial development was the extension of general equilibrium theory to dynamic stochastic models, with the simple yet important insight that commodities could be indexed not only by type, but also by date and event. This important insight was made by Arrow and Debreu (1954), who had important precursors in the work of Hicks (1939) and, particularly, in that of Lindahl (1929)—who had previously effectively extended competitive theory to dynamic environments. Subsequently, recursive methods, with their Markovian structure, were developed. These methods simplified the use of this dynamic framework and, in particular, its extension to stochastic general equilibrium analyses; see, for example, Stokey and Lucas (1989). Perhaps just as important as the development of tools for carrying out aggregate equilibrium analysis was the access to better and more systematic national income and product accounts data. In his review of growth theory, Solow (1970) lists the key growth facts which guided his research in growth theory in the 1950s. These growth facts were the relative constancy of investment and consumption shares of output, the relative constancy of labor and capital income shares, the continual growth of the real wage and output per capita, and the lack of trend in the return on capital. Solow (1956), in a seminal contribution, developed a simple model economy that accounted for these facts. The key to this early theory was the neoclassical production function, which is a part of the general equilibrium language. Afterwards the focus of attention shifted to preferences, with the important realization that the outcome of the CassKoopmans optimal growth model could be interpreted as the equilibrium of a competitive economy in which the typical consumer maximizes utility and the markets for both factors and products clear at every date. Scand. J. of Economics 1991 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 226 METHODOLOGY 168 F.E.Kydland and E.C.Prescott General Equilibrium Models By general equilibrium we mean a framework in which there is an explicit and consistent account of the household sector as well as the business sector. To answer some research questions, one must also include a sector for the government, which is subject to its own budget constraint. A model within this framework is specified in terms of the parameters that characterise preferences, technology, information structure, and institutional arrangements. It is these parameters that must be measured, and not some set of equations. The general equilibrium language has come to dominate in business cycle theory, as it did earlier in public finance, international trade, and growth. This framework is well-designed for providing quantitative answers to questions of interest to the business cycle student. One of these important questions, which has occupied business cycle theorists since the time of Frisch and Slutzky, is how to determine which sources of shocks give rise to cycles of the magnitudes we observe. To provide reliable answers to this and similar questions, abstractions are needed that describe the ability and willingness of agents to substitute commodities, both intertemporally and intratemporally, and within which one can bring to bear statistical or factual information. One of these abstractions is the neoclassical growth model. This model has proven useful in accounting for secular facts. To understand business cycles, we rely on the same ability and willingness of agents to substitute commodities as those used to explain the growth facts. We are now better able than Frisch was more than 50 years ago to calibrate the parameters of aggregate production technology. The wealth of studies on the growth model have shown us the way. To account for growth facts, it may be legitimate to abstract from the time allocation between market and nonmarket activities. To account for business cycle facts, however, the time allocation is crucial. Thus, good measures of the parameters of household technology are needed if applied business cycle theory is to provide reliable answers. The Econometrics of the General Equilibrium Approach The econometrics of the general equilibrium approach was first developed to analyze static or steady-state deterministic models. Pioneers of this approach are Johansen (1960) and Harberger (1962). This framework was greatly advanced by Shoven and Whalley (1972), who built on the work of Scarf (1973). Development was impeded by the requirement that there be a set of excess-demand functions, which are solved to find the equilibrium allocations. This necessitated that preference and technology structures have very special forms for which closed-form supply and demand functions existed. Perhaps these researchers were still under the influence of the systemof-equations approach and thought a model had to be a system of supply Scand. J. of Economics 1991 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE GENERAL EQUILIBRIUM APPROACH The econometrics of the general equilibrium approach 227 169 and demand functions. These researchers lacked the time series needed to estimate these equations. Given that they could not estimate the equations, they calibrated their model economy so that its static equilibrium reproduced the sectoral national income and product accounts for a base year. In their calibration, they used estimates of the elasticity parameters obtained in other studies. Their approach is ill-suited for the general equilibrium modeling of business fluctuations because dynamics and uncertainty are crucial to any model that attempts to study business cycles. To apply general equilibrium methods to the quantitative study of business cycle fluctuations, we need methods to compute the equilibrium processes of dynamic stochastic economies, and specific methods for the stochastic growth model economy. Recursive competitive theory and the use of linear-quadratic economies are methods that have proven particularly useful. These tools make it possible to compute the equilibrium stochastic processes of a rich class of model economies. The econometric problem arises in the selection of the model economies to be studied. Without some restrictions, virtually any linear stochastic process on the variables can be rationalized as the equilibrium behavior of some model economy in this class. The key econometric problem is to use statistical observations to select the parameters for an experimental economy. Once these parameters have been selected, the central part of the econometrics of the general equilibrium approach to business cycles is the computational experiment. This is the vehicle by which theory is made quantitative. The experiments should be carried out within a sensible or appropriate model economy that is capable of addressing the question whose answer is being sought. The main steps in econometric analyses are as follows: defining the question; setting up the model; calibrating the model; and reporting the findings. Question To begin with, the research question must be clearly defined. For example, in some of our own research we have looked at quantifying the contribution of changes in a technology parameter, also called Solow residuals, as a source of U.S. postwar business cycles. But we refined it further. The precise question asked is how much variation in aggregate economic activity would have remained if technology shocks were the only source of variation. We emphasize that an econometric, that is, quantitative theoretic analysis, can be judged only relative to its ability to address a clear-cut question. This is a common shortcoming of economic modeling. When the question is not made sufficiently clear, the model economy is often criticized for being ill-suited to answer a question it was never designed to answer. Scand. J. of Economics 1991 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 228 170 METHODOLOGY F.E.Kydland and E.C.Prescott Model Economy To address a specific question one typically needs a suitable model economy for addressing the specified question. In addition to having a clear bearing on the question, tractability and computability are essential in determining whether the model is suitable. Model-economy selection depends on the question being asked. Model-economy selection should not depend on the answer provided. Searching within some parametric class of economies for the one that best fits some set of aggregate time series makes little sense. Unlike the system-of-equations approach, no attempt is made to determine the true model. All model economies are abstractions and are by definition false. Calibration The model has to be calibrated. The necessary information can sometimes be obtained from data on individuals or households. An example of such information is the average fraction of discretionary time household members who are, or who potentially are, labor market participants actually spent in market activity. In many other cases, the required information easily can be obtained from aggregate nonbusiness-cycle information. The task often involves merely computing some simple averages, such as growth relations between aggregates. This is the case for inventory-output and capital-output ratios, and long-run fractions of the various G NP components to total output, among others. In some cases, history has provided sufficiently dramatic price experiments which can be used to determine, with a great deal of confidence, an elasticity of substitution. In the case of labor and capital as inputs in the aggregate business production function, and also in the case of consumption and leisure as inputs to household production, the large real-wage increase over several decades in relation to the prices of the other input, combined with knowledge about what has happened to the expenditure shares on the respective inputs, provides this kind of information. Because the language used in these business cycle models is the same as that used in other areas of applied economics, the values of common parameters should be identical across these areas and typically have been measured by researchers working in these other areas. One can argue that the econometrics of business cycles described here need not be restricted to general equilibrium models. In fact it is in the stage of calibration where the power of the general equilibrium approach shows up most forcefully. The insistence upon internal consistency implies that parsimoniously parameterized models of the household and business sector display rich dynamic behavior through the intertemporal substitution arising from capital accumulations and from other sources. Scand. J. of Economics 1991 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE GENERAL EQUILIBRIUM APPROACH The econometrics of the general equilibrium approach 229 171 Computational Experiments Once the model is calibrated, the next step is to carry out a set of computational experiments. If all the parameters can be calibrated with a great deal of accuracy, then only a few experiments are needed. In practice, however, a number of experiments are typically required in order to provide a sense of the degree of confidence in the answer to the question. It often happens that the answer to the research question is robust to sizable variations in some set of parameters and conclusions are sharp, even though there may be a great degree of uncertainty in those parameters. At other times, however, this is not the case, and without better measurement of the parameters involved, theory can only restrict the quantitative answer to a large interval. Findings The final step is to report the findings. This report should include a quantitative assessment of the precision with which the question has been answered. For the question mentioned above, the answer is a numerical estimate of the fraction of output variability that would have remained if variations in the growth of the Solow residual were the only source of aggregate fluctuation. The numerical answer to the research question, of course, is model dependent. The issue of how confident we are in the econometric answer is a subtle one which cannot be resolved by computing some measure of how well the model economy mimics historical data. The degree of confidence in the answer depends on the confidence that is placed in the economic theory being used. V. Two Applications to Business Cycle Theory We illustrate the econometrics of the general equilibrium approach to business cycle theory with two examples. The first example, credited to Lucas (1987) and Imrohoroglu (1989), addresses the question of quantifying the costs of business cycle fluctuations. An important feature of the quantitative general equilibrium approach is that it allows for explicit quantitative welfare statements, something which was generally not possible with the system-of-equations approach that preceded it. The second example investigates the question of how large business cycle fluctuations would have been if technology shocks were the only source of fluctuations. This question is also important from a policy point of view. If these shocks are quantitatively important, an implication of theory is that an important component of business cycle fluctuations is a good, not a bad. Scand. J. of Economics 1991 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 230 METHODOLOGY 172 F.E.Kydland and E.C.Prescott Costs of Business Cycle Fluctuations The economy Lucas uses for his quantitative evaluation is very simple. There is a representative or stand-in household and a random endowment process of the single consumption good. The utility function is standard, namely, the expected discounted value of a constant relative risk aversion utility function. Equilibrium behavior is simply to consume the endowment. Lucas determines how much consumption the agent is willing to forgo each period in return for the elimination of all fluctuations in consumption. Even with an extreme curvature parameter of 10, he finds that when the endowment process is calibrated to the U.S. consumption behavior, the cost per person of business cycle fluctuations is less than one-tenth of a per cent of per-capita consumption. This model abstracts from important features of reality. There is no investment good, and consequently no technology to transform the date t consumption good into the date t+1 consumption good. As the costs of fluctuation are a function of the variability in consumption and not in investment, abstracting from capital accumulation is appropriate relative to the research question asked. What matters for this evaluation is the nature of the equilibrium consumption process. Any representative-agent economy calibrated to this process will give the same answer to the question, so it makes sense to deal with the simplest economy whose equilibrium consumption process is the desired one. This is what Lucas does. Introducing the time-allocation decision between market and nonmarket activities would change the estimate, since the agent would have the opportunity to substitute between consumption and leisure. The introduction of these substitution opportunities would result in a reduction in the estimated cost of business cycle fluctuations as leisure moves countercyclically. But, given the small magnitude of the cost of business cycle fluctuations, even in a world without this substitution opportunity, and given that the introduction of this feature reduces the estimate of this cost, there is no need for its inclusion. In representative-agent economies, all agents are subject to the same fluctuations in consumption. If there is heterogeneity and all idiosyncratic risk is allocated efficiently, the results for the representative and heterogeneous agent economies coincide. This would not be the case if agents were to smooth consumption through the holding of liquid assets as is the case in the permanent income theory. Imrohoroglu (1989) examines whether the estimated costs of business cycle fluctuations are significantly increased if, as is in fact the case, people vary their holdings of liquid assets in order to smooth their stream of consumption. She modifies the Lucas economy by introducing heterogeneity and by giving each agent access to a technology that allows that agent to transform date t consumption into date t+1 consumption. Given that real interest rates were near zero in the Scand. J. of Economics 1991 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE GENERAL EQUILIBRIUM APPROACH The econometrics of the general equilibrium approach 231 173 fifty-odd years from 1933 to 1988, the nature of the storage technology chosen is that one unit of the good today can be transferred into one unit of the good tomorrow. She calibrates the processes on individual endowments to the per-capita consumption process, to the variability of annual income across individuals, and to the average holdings of the liquid asset—also across individuals. For her calibrated model economy, she finds the cost of business cycles is approximately three times as large as that obtained in worlds with perfect insurance of idiosyncratic risk. But three times a small number is still a small number. Technology Shocks as Source of Fluctuations One source of shocks suggested as far back as in Wicksell (1907) is fluctuations in technological growth. In the 1960s and 1970s, this source was dismissed by many as being unlikely to play much of a role in the aggregate. Most researchers accepted that there could be considerable variation in productivity at the industry level, but they believed that industry-level shocks would average out in the aggregate. During the 1980s, however, this source of shocks became the subject of renewed interest as a major source of fluctuations, in large part supported by quantitative economic theory. The question addressed, then, was how much would the U.S. postwar economy have fluctuated if technological shocks were the only source of aggregate fluctuations? Our selection of a model economy to address this question follows. First we extended the neoclassical growth model to include leisure as an argument of the stand-in household’s utility function. Given that more than half of business cycle fluctuations are accounted for by variations in the labor input, introducing this element is crucial. Next we calibrated the deterministic version of the model so that its consumption-investment shares, factor income shares, capital output ratios, leisure-market time shares, and depreciation shares matched the average values for the U.S. economy in the postwar period. Throughout this analysis, constant elasticity structures were used. As uncertainty is crucial to the question, computational considerations led us to select a linear-quadratic economy whose average behavior is the same as the calibrated deterministic constant elasticity of substitution economy. We abstracted from public finance considerations and consolidated the public and private sectors. We introduced Frisch’s (1933) assumption of time-to-build new productive capital. The construction period considered was four periods, with new capital becoming productive only upon completion, but with resources being used up throughout its construction. Given the high volatility of inventory investment, inventory stocks were included as a factor of production. We found, using the variance of Solow residuals estimated by Prescott (1986), that the model economy’s output Scand. J. of Economics 1991 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 232 174 METHODOLOGY F.E.Kydland and E.C.Prescott variance is 55 per cent as large as the corresponding variance for the U.S. economy in the postwar period. In the early 1980s, there was much discussion in the profession about the degree of aggregate intertemporal substitution of leisure. The feeling was that this elasticity had to be quite high in order for a market-clearing model to account for the highly volatile and procyclical movements in hours. This discussion may have started with the famous paper by Lucas and Rapping (1969). Realizing that the standard utility function implied a rather small elasticity of substitution, they suggested the possibility that past leisure choices may directly affect current utility. Being sympathetic to that view, we considered also a non-time-separable utility function as a tractable way of introducing this feature. When lags on leisure are considered, the estimate of how volatile the economy would have been if technology shocks were the only disturbance increases from 55 to near 70 per cent. But, until there is more empirical support for this alternative preference structure, we think estimates obtained using the economy with a time-separable utility function are better. Unlike the system-of-equations approach, the model economy which better fits the data is not the one used. Rather, currently established theory dictates which one is used. Probably the most questionable assumption of this theory, given the question addressed, is that of homogeneous workers, with the additional implication that all variation in hours occurs in the form of changes in hours per worker. According to aggregate data for the U.S. economy, only about one-third of the quarterly fluctuations in hours are of this form, while the remaining two-thirds arise from changes in the number of workers; see Kydland and Prescott (1989, Table 1). This observation led Hansen (1985) to introduce the Rogerson (1988) labor indivisibility construct into a business cycle model. In the Hansen world all fluctuations in hours are in the form of employment variation. To deal with the apparent nonconvexity arising from the assumption of indivisible labor, the problem is made convex by assuming that the commodity points are contracts in which every agent is paid the same amount whether that agent works or not, and a lottery randomly chooses who in fact works in every period. Hansen finds that with this labor indivisibility his model economy fluctuates as much as did the U.S. economy. Our view is that, with the extreme assumption of only fluctuations in employment, Hansen overestimates the amount of aggregate fluctuations accounted for by Solow residuals in the same way as our equally extreme assumption of only fluctuations in hours per worker lead us to an underestimation. In Kydland and Prescott (1989), the major improvement on the 1982 version of the model economy is to permit variation both in the number of workers and in the number of hours per worker. The number of hours a Scand. J. of Economics 1991 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE GENERAL EQUILIBRIUM APPROACH The econometrics of the general equilibrium approach 233 175 plant is operated in any given period is endogenous. The model also treats labor as a quasi-fixed input factor by assuming costs of moving people into and out of the business sector. Thus, in this model there is what we interpret to be labor hoarding. Without the cost of moving workers in and out of the labor force, a property of the equilibrium turns out to be that all the hours variation is in the form of employment change and none in hours per worker. In that respect, it is similar to Hansen’s (1985) model. For this economy with no moving costs, the estimate is that Solow residuals account for about 90 per cent of the aggregate output variance. For this economy with moving costs, we calibrated so that the relative variations in hours per worker and number of workers matched U.S. data. With this degree of labor hoarding, the estimate of the fraction of the cycle accounted for by Solow residuals is reduced to 70 per cent. A widespread and misguided criticism of our econometric studies, for example, McCallum (1989), is that the correlation between labor productivity and the labor input is almost one for our model economy while it is approximately zero for the U.S. postwar economy. If we had found that technology shocks account for nearly all fluctuations and that other factors were unimportant, the failure of the model economy to mimic the data in this respect would cast serious doubt on our findings. But we did not find that the Solow technology shocks are all-important. We estimate that these technology shocks account for about 70 per cent of business cycle fluctuations. If technology shocks account for 70 per cent, and some other shocks which are orthogonal to technology shocks account for 30 per cent, theory implies a correlation between labor productivity and the labor input near zero. Christiano and Eichenbaum (1990) have established this formally in the case that the other shock is variations in public consumption. But the result holds for any shock that is orthogonal to the Solow technology shocks. The fact that this correlation for our model economy and the actual data differ in the way they do adds to our confidence in our findings. The estimate of the contribution of technology shocks to aggregate shocks has been found to be robust to several modifications in the model economy. For example, Greenwood, Hercowitz, and Huffman (1988) permit the utilization rate of capital to vary and to affect its depreciation rate, while all technology change is embodied in new capital; Danthine and Donaldson (1989) introduce an efficient-wage construct; Cooley and Hansen (1989) consider a monetary economy with a cash-in-advance constraint; and RiosRull (1990) uses a model calibrated to life cycle earnings and consumption patterns. King, Plosser, and Rebelo (1988) have non-zero growth. Gomme and Greenwood (1990) have heterogenous agents with recursive preferences and equilibrium risk allocations. Benhabib, Rogerson, and Wright (1990) incorporate home production. Hornstein (1990) considers the implications Scand. J. of Economics 1991 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 234 176 METHODOLOGY F.E.Kydland and E.C.Prescott of increasing returns and monopolistic competition. In none of these cases is the estimate of the contribution of technology shocks to aggregate fluctuations significantly altered. VI. Concluding Remarks Econometrics is by definition quantitative economic theory—that is, economic analyses which provide quantitative answers to clear-cut questions. The general equilibrium econometric methodology is centered around computational experiments. These experiments provide answers to the questions posed in the model economies whose equilibrium elements have been computed. The model economy selected should quantitatively capture people’s ability and willingness to substitute and the arrangements employed which are relevant to the question. We base our quantitative economic intuition on the outcome of these experiments. The dramatic advances in econometric methodology over the last 25 years have made it possible to apply fully neoclassical econometrics to the study of business cycles. Already there have been several surprising findings. Contrary to what virtually everyone thought, including the authors of this review, technology shocks were found to be an important contributor to business cycle fluctuations in the U.S. postwar period. Not all fluctuations are accounted for by technology shocks, and monetary shocks are a leading candidate to account for a significant fraction of the unaccounted-for aggregate fluctuations. The issue of how to incorporate monetary and credit factors into the structure is still open, with different avenues under exploration. When there is an established monetary theory, we are sure that general equilibrium methods will be used econometrically to evaluate alternative monetary and credit arrangements. References Arrow, Kenneth J. & Debreu, Gerard: Existence of an equilibrium for a competitive economy. Econometrica 22(3), 265–90, 1954. Benhabib, Jess, Rogerson, Richard & Wright, Randall: Homework in Macroeconomics I: Basic Theory, mimeo, 1990. Brunner, Karl: The disarray in macroeconomics. In Forrest Capie & Geoffrey E.Wood (eds.), Monetary Economics in the 1980s, MacMillan Press, New York, 1989. Christiano, Lawrence J. & Eichenbaum, Martin: Current real business cycle theories and aggregate labor market fluctuations. DP 24, Institute for Empirical Macroeconomics, Federal Reserve Bank of Minneapolis and University of Minnesota, 1990. Cooley, Thomas F. & Hansen, Gary D.: The inflation tax in a real business cycle model. American Economic Review 79(4), 733–48, 1989. Scand. J. of Economics 1991 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE GENERAL EQUILIBRIUM APPROACH The econometrics of the general equilibrium approach 235 177 Danthine, Jean-Pierre & Donaldson, John B.: Efficiency wages and the real business cycle. Cahier 8803, Départment d’économétri et d’économie politique, Université de Lausanne, 1988; forthcoming in European Economic Review. Frisch, Ragnar: Propagation problems and impulse problems in dynamic economics. In Economic Essays in Honor of Gustav Cassel, London, 1933. Frisch, Ragnar: Econometrics in the world of today. In W.A.Eltis, M.F.G.Scott & J.N. Wolfe (eds.), Induction, Growth and Trade: Essays in Honour of Sir Roy Harrod, Clarendon Press, Oxford, 152–66, 1970. Greenwood, Jeremy, Hercowitz, Zvi & Huffman, Gregory W.: Investment, capacity utilization and the business cycle. American Economic Review 78, 402–18, June 1988. Gomme, Paul & Greenwood, Jeremy: On the cyclical allocation of risk. WP 462, Research Department, Federal Reserve Bank of Minneapolis, 1990. Harberger, Arnold C: The incidence of the corporation income tax. Journal of Political Economy 70(3), 215–40, 1962. Haberler, Gottfried: “Discussion” of the econometric approach to business fluctuations by Tjalling C.Koopmans. American Economic Review 39, 84–8, May 1949. Hansen, Gary D.: Indivisible labor and the business cycle. Journal of Monetary Economics 16(3), 309–27, 1985. Hicks, John R.: Value and Capital: An Inquiry into Some Fundamental Principles of Economic Theory. Clarendon Press, Oxford, 1939. Hornstein, Andreas: Monopolistic competition, increasing returns to scale, and the importance of productivity changes. WP, University of Western Ontario, 1990. Imrohoroglu, Ayse: Costs of business cycles with indivisibilities and liquidity constraints. Journal of Political Economy 97, 1364–83, Dec. 1989. Johansen, Leif: A Multi-sectoral Study of Economic Growth. North-Holland, Amsterdam, 1960. King, Robert G., Plosser, Charles I. & Rebelo, Sergio T.: Production, growth and business cycles II: New directions. Journal of Monetary Economics 21, 309–41, March/May 1988. Koopmans, Tjalling C.: The econometric approach to business fluctuations. American Economic Review 39, 64–72, May 1949. Kydland, Finn E. & Prescott, Edward C.: Time to build and aggregate fluctuations. Econometrica 50, 1345–70, 1982. Kydland, Finn E. & Prescott, Edward C.: Hours and employment variation in business cycle theory. DP 17, Institute for Empirical Macroeconomics, Federal Reserve Bank of Minneapolis and University of Minnesota, 1989; forthcoming in Economic Theory. Kydland, Finn E. & Prescott, Edward C.: Business cycles: Real facts and a monetary myth. Federal Reserve Bank of Minneapolis Quarterly Review, 3–18, Spring 1990. Lindahl, Erik: Prisbildningsproblemets uppläggning från kapitalteoretisk synpunkt. Ekonomisk Tidskrift 31, 31–81, 1929. Translated as: The place of capital in the theory of price, in Studies in the Theory of Money and Capital, Farrar and Reinhart, New York, 269–350, 1929. Lucas, Robert E., Jr.: Models of Business Cycles. Yrjö Jahnsson Lectures, Basil Blackwell, Oxford and New York, 1987. Scand. J. of Economics 1991 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 236 178 METHODOLOGY F.E.Kydland and E.C.Prescott Lucas, Robert E., Jr. & Rapping, Leonard A.: Real wages, employment and inflation. Journal of Political Economy 77, 721–54, Sept./Oct. 1969. Lucas, Robert E., Jr. & Sargent, Thomas J.: After Keynesian macroeconomics. In After the Phillips Curve: Persistence of High Inflation and High Unemployment, Conference Series No. 19, Federal Reserve Bank of Boston, 49–72, 1978. McCallum, Bennett T.: Real business cycle models. In R.J.Barro (ed.), Modern Business Cycle Theories, Harvard University Press, Boston, 16–50, 1989. Prescott, Edward C.: Theory ahead of business cycle measurement. CarnegieRochester Conference Series on Public Policy 25, 11–44, 1986. Rios-Rull, Jose Victor: Life cycle economies and aggregate fluctuations. Preliminary draft, Carnegie-Mellon University, June 1990. Rogerson, Richard: Indivisible labor, lotteries and equilibrium. Journal of Monetary Economics 21, 3–16, Jan. 1988. Samuelson, Paul A.: Interaction between the multiplier analysis and the principle of acceleration. Review of Economic and Statistics 29, 75–8, May 1939. Scarf, Herbert (with the collaboration of T.Hansen): Computation of Economic Equilibria. Yale University Press, New Haven, 1973. Schumpeter, Joseph: The common sense of econometrics. Econometrica 1, 5– 12, Jan. 1933. Shoven, John B. & Whalley, John: A general equilibrium calculation of the effects of differential taxation of income from capital in the U.S. Journal of Public Economics 1 (3/4), 281–321, 1972. Solow, Robert M.: A contribution to the theory of economic growth. Quarterly Journal of Economics 70 (1), 65–94, 1956. Solow, Robert M.: Technical change and the aggregate production function. Review of Economics and Statistics 39(3), 312–20, 1957. Solow, Robert M.: Growth Theory: An Exposition. Radcliffe Lectures, Clarendon Press, Oxford, 1970. Stokey, Nancy & Lucas, Robert E., Jr., with Prescott, Edward C.: Recursive Methods in Economic Dynamics. Harvard University Press, Cambridge, MA, 1989. Tinbergen, Jan: Annual survey: Suggestions on quantitative business cycle theory. Econometrica 3, 241–308, July 1935. Wicksell, Knut: Krisernas gåta. Statsøkonomisk Tidsskrift 21, 255–70, 1907. Scand. J. of Economics 1991 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors CHAPTER 13 237 Journal of Economic Perspectives—Volume 10, Number 1—Winter 1996—Pages 69–85 The Computational Experiment: An Econometric Tool Finn E.Kydland and Edward C.Prescott I n a computational experiment, the researcher starts by posing a welldefined quantitative question. Then the researcher uses both theory and measurement to construct a model economy that is a computer representation of a national economy. A model economy consists of households, firms and often a government. The people in the model economy make economic decisions that correspond to those of their counterparts in the real world. Households, for example, make consumption and savings decisions, and they decide how much to work in the market. The researcher then calibrates the model economy so that it mimics the world along a carefully specified set of dimensions. Finally, the computer is used to run experiments that answer the question.1 Such experiments have become invaluable tools in quantitative aggregate theory.2 They are being used, for example, to estimate the quantitative effects of trade liberalization policies, measure the welfare consequences of changes in the tax system and quantify the magnitude and nature of business cycle fluctuations induced by different types of shocks. In this paper, we review the use of the computational experiment in economics. 1 Lucas (1980), in his paper on methods and problems in business cycle theory, explains the need for computational experiments in business cycle research. 2 Shoven and Whalley (1972) were the first to use what we call the computational experiment in economics. The model economies that they used in their experiments are static and have many industrial sectors. Finn E.Kydland is Professor of Economics, Carnegie-Mellon University, Pittsburgh, Pennsylvania, and Research Associate, Federal Reserve Bank of Cleveland, Cleveland, Ohio. Edward C.Prescott is Professor of Economics, University of Minnesota, and Research Associate, Federal Reserve Bank of Minneapolis, both in Minneapolis, Minnesota. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 238 METHODOLOGY 70 Journal of Economic Perspectives One immediate question that arises is whether the computational experiment should be regarded as an econometric tool (for example, Gregory and Smith, 1993). In the modern (narrow) sense of the term it is not, since it isn’t used in the “measurement of economic relations” (Marschak, 1948, p. 1). Yet it is an econometric tool in the original (broad) sense of the term (which we prefer), since computational experiments are used to derive the quantitative implications of economic theory (Frisch, 1933a, p. 1). In Kydland and Prescott (1991a), we develop the position that the computational experiment is an econometric tool, but here, we avoid this largely semantic debate. Instead, we will simply state that the task of deriving the quantitative implications of theory differs from that of measuring economic parameters. Computational experiments are not unique to economic science—they are heavily used in the physical sciences as well. In one crucial respect, however, they do differ across the two disciplines. Unlike theory in the physical sciences, theory in economics does not provide a law of motion. Rather, economic theory provides a specification of people’s ability and willingness to substitute among commodities. Consequently, computational experiments in economics include the additional step of computing the equilibrium process in which all of the model’s people behave in a way that is in each person’s best interest—that is, economists must compute the equilibrium law of motion or process of the model economy. Given the process governing the system, both economic and physical science use the computer to generate realizations of this process. If the model is deterministic, only one possible equilibrium realization exists for the path of the model economy. If the model economy has aggregate uncertainty—as it must, for example, if the phenomena of interest are business cycle fluctuations—then the model will imply a process governing the random evolution of the economy. In the case of uncertainty, the computer can generate any number of independent realizations of the equilibrium stochastic process, and these realizations, along with statistical estimation theory, are then used to measure the sampling distribution of any desired set of statistics of the model economy. Steps in an Economic Computational Experiment Any economic computational experiment involves five major steps: pose a question; use a well-tested theory; construct a model economy; calibrate the model economy; and run the experiment. We will discuss each of these steps in turn. Pose a Question The purpose of a computational experiment is to derive a quantitative answer to some well-posed question. Thus, the first step in carrying out a computational experiment is to pose such a question. Some of these questions are concerned with policy evaluation issues. These questions typically ask about the welfare and distributive consequences of some policy under consideration. Other © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE COMPUTATIONAL EXPERIMENT Finn E.Kydland and Edward C.Prescott 239 71 Table 1 Examples of Well-Posed Questions in Studies Using the Computational Experiment questions are concerned with the testing and development of theory. These questions typically ask about the quantitative implications of theory for some phenomena. If the answer to these questions is that the predictions of theory match the observations, theory has passed that particular test. If the answer is that there is a discrepancy, a deviation from theory has been documented. Still, other experiments are concerned with the sensitivity of previous findings to the introduction of some feature of reality from which previous studies have abstracted. Table 1 offers some examples of computational experiments that seek to answer each of these types of questions. That table highlights the fact that judging whether a model economy is a “good” abstraction can be done only relative to the question posed. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 240 METHODOLOGY 72 Journal of Economic Perspectives Use Well-Tested Theory With a particular question in mind, a researcher needs some strong theory to carry out a computational experiment: that is, a researcher needs a theory that has been tested through use and found to provide reliable answers to a class of questions. Here, by theory we do not mean a set of assertions about the actual economy. Rather, following Lucas (1980), economic theory is defined to be “an explicit set of instructions for building…a mechanical imitation system” to answer a question. If the question is quantitative in nature, a computer representation of the imitation system or economy is used, and extensive computations are required to answer the posed question. As one example, the computational experiments often carried out in modern business cycle theory build upon the neoclassical growth framework. Central to neoclassical growth theory is its use of an aggregate production function, with the output of goods resulting from inputs of labor and capital.3 This framework has served well when dealing with growth within reasonably stable economic institutions. With an explicit description of the household sector, including its focus on the time-allocation decision, the neoclassical growth model becomes an internally consistent framework for addressing business cycle questions, as well as other questions of interest to macroeconomists. The theory implies that when a model economy is confronted with technology, public finance and terms-of-trade shocks, it should display business cycle fluctuations of a quantitative nature similar to those actually observed. In other words, modern business cycle models are stochastic versions of neoclassical growth theory. And the fact that business cycle models do produce normal-looking fluctuations adds dramatically to our confidence in the neoclassical growth theory model—including the answers it provides to growth accounting and public finance questions. We recognize, of course, that although the economist should choose a welltested theory, every theory has some issues and questions that it does not address well. In the case of neoclassical growth theory, for example, it fails spectacularly when used to address economic development issues. Differences in stocks of reproducible capital stocks cannot account for international differences in per capita incomes. This does not preclude its usefulness in evaluating tax policies and in business cycle research. Construct a Model Economy With a particular theoretical framework in mind, the third step in conducting a computational experiment is to construct a model economy. Here, key issues are the amount of detail and the feasibility of computing the equilibrium process. Of-ten, economic experimenters are constrained to deal with a much 3 Neoclassical growth theory also represents a good example of the importance of interaction between factual studies and theory development. Solow (1970) lists several growth facts that influenced the development of neoclassical growth theory. Once the main ingredients of the theory were established—such as the production function—new light was thrown on the data. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE COMPUTATIONAL EXPERIMENT 241 The Computational Experiment: An Econometric Tool 73 simpler model economy than they would like because computing the equilibrium of a more complicated model would be impossible, given currently available tools. This situation is no different from that in the physical sciences, where, as in economics, the computational experiment has become accepted as an invaluable scientific tool. For example, in his overview of climate modeling, Schneider (1987, p. 72) states: “Although all climate models consist of mathematical representations of physical processes, the precise composition of a model and its complexity depend on the problem it is designed to address.” And later (p. 72): “Often it makes sense to attack a problem first with a simple model and then employ the results to guide research at higher resolution.” In the physical sciences, as in economics, confidence in a particular framework or approach is gained through successful use. So far, most of the model environments that macroeconomists have used share certain characteristics. The environments are inhabited by a large number of people whose decision problems are described explicitly. Both the household and business sectors play a central role. For some questions, government or foreign sectors must be included as well. In some models everyone is alike; in others, such as those designed to address questions where demographic changes are important, heterogeneous people must be used. This description may sound somewhat indefinite or abstract, but we reemphasize that an abstraction can be judged only relative to some given question. To criticize or reject a model because it is an abstraction is foolish: all models are necessarily abstractions. A model environment must be selected based on the question being addressed. For example, heterogeneity of people is crucial in the Auerbach and Kotlikoff (1987) model, which predicts the consequences of the population’s changing age distribution on savings. However, Ríos-Rull (1994) demonstrates that such life cycle features, even when combined with elements of market incompleteness, are not quantitatively important to business cycle findings regarding issues such as the contribution of technology shocks to business cycle fluctuations. The features of a given model may be appropriate for some question (or class of questions) but not for others.4 The selection and construction of a particular model economy should not depend on the answer provided. In fact, searching within some parametric class of economies for the one that best fits a set of aggregate time series makes little sense, because it isn’t likely to answer an interesting question. For example, if the question is of the type, “how much of fact X is accounted for 4 We will not debate the legitimacy of these methods, since such debates generally serve to define schools rather than to produce agreement. Such debates are almost nonexistent during normal science, but tend to recur during scientific revolutions. As stated by Kuhn (1962, p. 145), “Few philosophers of science still seek absolute criteria for the verification of scientific theories.” All historically significant theories have agreed with the facts, but only to a degree. No more precise answer can be found to the question of how well an individual theory fits the facts. Using probabilistic verification theories that ask us to compare a given scientific theory with all others that might fit the same data is a futile effort. We agree with Kuhn (p. 146) that “probabilistic theories disguise the verification situation as much as they illuminate it.” © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 242 METHODOLOGY 74 Journal of Economic Perspectives by Y,” then choosing the parameter values in such a way as to make the amount accounted for as large as possible according to some metric is an attempt to get a particular—not a good—answer to the question. Calibrate the Model Economy Now that a model has been constructed, the fourth step in carrying out a computational experiment is to calibrate that model. Originally, in the physical sciences, calibration referred to the graduation of measuring instruments. For example, a Celsius thermometer is calibrated to register zero degrees when immersed in water that contains ice and 100 degrees when immersed in boiling water. A thermometer relies on the theory that mercury expands (approximately) linearly within this range of temperatures. Related theory also tells us how to recalibrate the thermometer if the measurements are made in Denver or Mexico City rather than at sea level. In a sense, model economies, like thermometers, are measuring devices. Generally, some economic questions have known answers, and the model should give an approximately correct answer to them if we are to have any confidence in the answer given to the question with unknown answer. Thus, data are used to calibrate the model economy so that it mimics the world as closely as possible along a limited, but clearly specified, number of dimensions. Note that calibration is not an attempt at assessing the size of something: it is not estimation. Estimation is the determination of the approximate quantity of something. To estimate a parameter, for example, a researcher looks for a situation in which the signal-to-noise ratio is high. Using the existing data and some theory, the researcher then constructs a probability model. An estimator is developed that is robust, relative to the parameter that is to be estimated, to the questionable features of the maintained hypothesis. As a second example of estimation, a computational experiment itself is a type of estimation, in the sense that the quantitative answer to a posed question is an estimate. For example, the quantitative theory of a computational experiment can be used to measure the welfare implications of alternative tax policies. It is important to emphasize that the parameter values selected are not the ones that provide the best fit in some statistical sense. In some cases, the presence of a particular discrepancy between the data and the model economy is a test of the theory being used. In these cases, absence of that discrepancy is grounds to reject the use of the theory to address the question. One such example is the use of standard theory to answer the question of how volatile the postwar U.S. economy would have been if technology shocks had been the only contributor to business cycle fluctuations. To answer this question, a model economy with only technology shocks was needed. Using the standard neoclassical production function, standard preferences to describe people’s willingness to substitute intra- and intertemporally between consumption and leisure, and an estimate of the technology shock variance, we found that the model economy displays business cycle fluctuations 70 percent as large as did the U.S. economy (Kydland and Prescott, 1991b). This number is our answer to the posed question. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE COMPUTATIONAL EXPERIMENT Finn E.Kydland and Edward C.Prescott 243 75 Some have questioned our finding, pointing out that on one key dimension real business cycle models and the world differ dramatically: the correlation between hours worked and average labor productivity is near one in the model economy and approximately zero in U.S. postwar observations (McCallum, 1989). The detractors of the use of standard theory to study business cycles are correct in arguing that the magnitude of this correlation in the world provides a test of the theory. They are incorrect in arguing that passing this test requires the value of this correlation in the model and in the real world to be approximately equal. An implication of the theory is that this correlation is a function of the importance of technology shocks relative to other shocks. In particular, the less is the relative importance of technology shocks, the smaller this correlation should be. The reason for this is that the factors other than technology shocks that give rise to variation in the labor input result in productivity being low when hours are high.5 Given that the estimated contribution of technology shocks to fluctuations is 70 percent, the correlation between hours and labor productivity being near one in the data would have been grounds for dismissing our answer. (For further elaboration on this point, see Kydland and Prescott, 1991b, p. 79; Aiyagari, 1994.) Run the Experiment The fifth and final step in conducting a computational experiment is to run the experiment. Quantitative economic theory uses theory and measurement to estimate how big something is. The instrument is a computer program that determines the equilibrium process of the model economy and uses this equilibrium process to generate equilibrium realizations of the model economy. The computational experiment, then, is the act of using this instrument. These equilibrium realizations of the model economy can be compared with the behavior of the actual economy in some period as follows. If the model economy has no aggregate uncertainty, then it is simply a matter of comparing the equilibrium path of the model economy with the path of the actual economy. If the model economy has aggregate uncertainty, first a set of statistics that summarize relevant aspects of the behavior of the actual economy is selected. Then the computational experiment is used to generate many independent realizations of the equilibrium process for the model economy. In this way, the sampling distribution of this set of statistics can be determined to any degree of accuracy for the model economy and compared with the values of the set of statistics for the actual economy. In comparing the sampling distribution of a statistic for the model economy to the value of that statistic for the actual data, it is crucial that the same statistic be computed for the model and the real world. If, for example, the statistic for 5 Christiano and Eichenbaum (1992a) have established formally this possibility in the case where the other shock is variations in public consumption, but the result holds for any shock that is approximately orthogonal to the technology shocks. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 244 METHODOLOGY 76 Journal of Economic Perspectives the real world is for a 50-year period, then the statistic for the model economy must also be for a 50-year period. The Computational Experiment in Business Cycle Research Business cycles, that is, the recurrent fluctuations of output and employment about trend, puzzled economists for a long time. Understanding business cycles required the development of methods that made possible the use of computational experiments to answer questions concerned with the behavior of dynamic economies with uncertainty. Prior to the development of these methods, business cycle fluctuations were viewed as deviations from theory, and very little progress was made in understanding them. Subsequent to the development of those methods, computational experiments have been extensively used in business cycle research. The results of these experiments forced economists to revise their old views, and business cycle fluctuations are now seen as being, in fact, predicted by standard theory. For these reasons, we choose business cycle theory to illustrate the use of computational experiments in economic research. Posing Questions about the Business Cycle In the 1970s, a common assumption behind research on the business cycle was that one set of factors, most likely monetary shocks, was behind the cyclical component and that an entirely different set of factors, mainly the growth of productivity and inputs summarized by the neoclassical growth model, accounted for the movement of the long-run growth component. But there was also an earlier view, tracing as far back as work by Wicksell (1907), that suggested that fluctuation in technological growth could produce broad economic fluctuations. In the 1960s and 1970s, this source was dismissed by many as being unlikely to play much of a role in the aggregate. Most researchers believed that considerable variation could exist in productivity at the industry level, but they believed that industry-level shocks would average out in the aggregate. During the 1980s, however, technology shocks gained renewed interest as a major source of fluctuations, supported largely by computational experiments and quantitative economic theory. As a consequence, business cycle theory treats growth and cycles as being integrated, not as a sum of two components driven by different factors.6 6 An operational way of defining the trend empirically is described in Hodrick and Prescott (1980), who used standard curve-fitting techniques to define a growth component as being the curve that best fits a time series in a least-square sense, subject to a penalty on the sum of the second differences squared. The larger this penalty parameter, the smoother the fitted curve. For quarterly series, they found that a penalty parameter of 1600 made the fitted curve mimic well the one that business cycle analysts would draw. Given the finding that business cycle fluctuations are quantitatively just what neoclassical growth theory predicts, the resulting deviations from trend are nothing more than well-defined statistics. We emphasize that given the way the theory has developed, these statistics measure nothing. Business cycle theory treats growth and cycles as being integrated, not as a sum of © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE COMPUTATIONAL EXPERIMENT 245 The Computational Experiment: An Econometric Tool 77 Thus, the fundamental question here is the extent to which neoclassical growth theory can account for business cycle fluctuations, as well as long-term growth trends. A particular question addressed was, “How much would the U.S. postwar economy have fluctuated if technology shocks had been the only source of fluctuations?” Computational experiments are well suited to tackle this question. The Theory Used in Model Selection The basic theory used in the modern study of business cycles is the neoclassical growth model. The basic version of this model can best be understood as based on five relationships. The first relationship is an aggregate production function that sets total output equal to AtF(Kt, Ht), where F is a constant returns to scale function where the inputs are capital and labor, and At is the technology level that grows at random rates. In the simplest case, aggregate output is divided between consumption C and investment I. Under the assumption that factors are paid their marginal product, we obtain the identity that GNP and income are equal: C+I=wH+rK, where w and r are factor rental prices. The second relationship in the model economy describes the evolution of the capital stock. Each time period, the existing capital stock depreciates at a constant rate δ, but is replenished by new investment It. Thus Kt+1=(1–δ)Kt+It. The third relationship describes the evolution of the technology parameter At. Given that a structure that displays persistence is needed, a typical form would be At+1=ρAt+⑀t+1, where ρ is large but less than one, and the shocks ⑀t+1 are identically and independently distributed. In other words, the technology level for any given period depends on the technology level in the previous period, plus a random disturbance. The technology described by these relations specifies people’s ability to substitute. The fourth relationship needed for a fully specified economy is a specification of people’s willingness to substitute between consumption and leisure, both intertemporally and intratemporally. For this purpose, our model economy has a stand-in household with utility function that depends on consumption and leisure. 7 For simplicity, one can assume that the households own the capital stock directly and rent it to the firms. two components driven by different factors. For that reason, talking about the resulting statistics as imposing spurious cycles makes no sense. The Hodrick-Prescott filter is simply a statistical decomposition that summarizes in a reasonable way what happens at business cycle frequencies. The representation has proven useful in presenting the findings and judging the reliability of the answer, as well as a way of demonstrating remaining puzzles or anomalies relative to theory. 7 More explicitly, the function can take the general form where we normalize so that market and nonmarket productive time add to one. For a complete specification of the model, values of the parameters β, δ and ρ are needed, as well as the explicit utility function U and the production function F. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 246 METHODOLOGY 78 Journal of Economic Perspectives The final required element is an equilibrium concept. The one used is the competitive equilibrium, which equates marginal rates of substitution and transformation to price ratios. This involves equilibrium decision functions for consumption, investment and labor input as functions of the capital stock and productivity level during that time period: C(Kt, At), I(Kt, At) and H(Kt, At), respectively. In other words, using dynamic programming methods, the decisions can be computed as functions of the list of state variables that provide sufficient information about the position of the economy. Through this theory, business cycle theorists make contact with other fields of economics. Macroeconomics is no longer largely separate from the rest of economics. The utility and production functions used by business cycle theorists are similar to those used by public finance researchers (for example, Auerbach and Kotlikoff, 1987). The introduction of household production illustrates the close connection with the work of labor economists (for example, Benhabib, Rogerson and Wright, 1991; Greenwood and Hercowitz, 1991). When these models are expanded to consider international trade explicitly, they draw upon work in that field (Backus, Kehoe and Kydland, 1994). The choice of a model, as already noted, must be governed both by the question at hand and by what is computable. As an example of altering the model to suit the posed question, consider a contrast between two otherwise similar models. Benhabib, Rogerson and Wright (1991) and Greenwood and Hercowitz (1991) both consider household production in addition to market production, but the two studies are motivated by somewhat different questions. Both use capital and labor as inputs in nonmarket production. Benhabib and his coauthors divide the time allocation of households into three uses: market and nonmarket production time and leisure time. The model is designed to capture the household decision to combine its labor with machines, such as stoves and washing machines, to produce household consumption services. The authors argue that houses do not need to be combined with labor, at least not to the same extent that household machines do, so they exclude housing capital from their concept of household capital. Greenwood and Hercowitz, on the other hand, distinguish only between market and nonmarket time and include the stock of housing, along with consumer durables, in their concept of household capital.8 This example illustrates the important point that even the definition of particular variables in relation to the model economy may depend on the question. If a model environment is not computable, then it cannot be used for a computational experiment. This restriction can be a severe one, and the development of appropriate computable methods must therefore be given high priority. Fortunately, there has been considerable progress in this area over the last 30 or 40 years. In cases involving uncertain intertemporal behavior, the development of statistical decision theory has provided a 8 To be consistent, they then subtract gross housing product (the measure of the service flow from the economy’s housing stock) from GNP and add it to the consumer durables component of personal consumption expenditures. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE COMPUTATIONAL EXPERIMENT Finn E.Kydland and Edward C.Prescott 247 79 consistent way for people to make decisions under uncertainty. Another significant development is the Arrow-Debreu general equilibrium theory, which extends equilibrium theory to uncertain environments.9 More recently, Ríos-Rull (1995) offers an overview of the expansion in computable general equilibrium models that incorporate heterogeneity in the household sector—a category that has expanded dramatically over the last few years. Calibration Often, calibration involves the simple task of computing a few averages. For example, if the standard Cobb-Douglas production function is used—that is, if we let F(K, H) = K1-θHθ—then a numerical value for the parameter θ can be obtained by computing the average labor share of total output over a period of years. Several other growth relations map more or less directly into parameter values for typical models within the neoclassical growth framework, at least if the functional forms have been chosen with calibration in mind. Most growth relations have not changed much, on average, from one cycle to the next for several decades. As a consequence, computational experiments replicate the key long-term or growth relations among model aggregates. Exceptions do exist, where long-term relationships are not stable. For example, the inventory stock as a fraction of G NP has declined steadily. Durables expenditures as a fraction of total output have risen. For some purposes these changes can be ignored, since that feature does not significantly affect the answer to the question posed. At other times, depending on the associated pattern in the corresponding relative price, such information enables the researcher to obtain a quite precise estimate of some elasticity of substitution, which can then be built into the model. A good example of this sort of issue is the fact that hours of work per household are about the same now as four decades ago, in spite of a large rise in the real wage rate over the same period. This fact indicates that the elasticity of substitution between consumption and nonmarket time is near one. Still, many business cycle models abstract from the long-run productivity growth that is required to imply this sort of wage growth, because the answer to the questions addressed in those studies would have been essentially the same, as shown by Hansen (1986).10 To calibrate a utility function for the household sector of the economy, it is common to rely on averages across large numbers of the relevant population in the actual economy. For example, some model environments employ a utility function in consumption and leisure that, like the Cobb-Douglas 9 Also important is the development of recursive methods for the study of economic dynamics, because these methods allow economists to use the computational experiment to generate time series disciplined by factual studies (Stokey and Lucas, 1989). 10 Hansen (1986) compares otherwise identical model economies and permits growth in one version and not in the other. The model without growth needs a slight adjustment in the capital depreciation rate to be calibrated to the investment share of output and the observed capital/output ratio. With this adjustment, both models estimate the same role of technological shocks (more precisely, Solow residuals) for cyclical fluctuations. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 248 80 METHODOLOGY Journal of Economic Perspectives production function above, has a share parameter. In this case, the weight that should be placed on consumption turns out to be approximately equal to the average fraction of time spent in market activity. This fraction, in principle, can be obtained from panel data for large samples of individuals. Ghez and Becker (1975) offer a careful measurement study—making reasonable and thoughtful judgments about factors like age limits of the population sample and definition of total time allocated to market and nonmarket activities, including treatment of sleep and personal care. In calibration, we sometimes make the model economy inconsistent with the data on one dimension so that it will be consistent on another. For example, Imrohoroglu (1992) explores the welfare consequences of alternative monetary arrangements in worlds where agents are liquidity constrained, while Cooley and Hansen (1989) explore the welfare consequences in worlds where people use money for transaction purposes. These are two very different environments, each of which abstracts from the main feature of the other. Imrohoroglu calibrates her model economy to yield a stock of money held per household that is in line with U.S. observations. In her model, however, people hold money because they do not have access to an insurance technology to insure against randomness in the market value of their time. Equivalently, if they do have access to such an insurance technology, they find it so costly that, in equilibrium, they do not employ it. This is the only reason, in her model, for people to hold money; if she had calibrated the model to the amount of variation in individual income found in panel data, the model would have implied that average household holdings of liquid assets were about half of those actually held. Of course, households have other reasons for holding liquid assets that earn much less than the average return on physical capital. For instance, such assets can be used as a down payment on a house at some future date, as a substitute for insurance against sickness and accidents, or for transaction purposes, as in the Cooley and Hansen (1989) environment. These and other factors are abstracted from in the world of Imrohoroglu (1992), which led her to introduce greater variation in the market value of households’ time so as to make per capita holdings of money in the model match actual holdings. This calibration is reasonable, given the question she addresses. Her implicit assumption is that it is unimportant which liquidity factor gives rise to these holdings. Subsequent research will either support this working hypothesis or disprove it and, in the process, lead to better model economies for evaluating monetary and credit policy arrangements. This sequence is how economic science progresses. Running Experiments With explicit functional forms for the production and utility functions, with values assigned to the parameters, and with a probability distribution for the shocks, a researcher can use the model economy to perform computational experiments. The first step is to compute the equilibrium decision rules, which are functions of the state of the economy. The next step is to generate equilibrium realizations of the economy. The computer begins each period © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE COMPUTATIONAL EXPERIMENT 249 The Computational Experiment: An Econometric Tool 81 with a given level of the state variables, for example, the capital stock and the technology level. The values of the state variables along with the equilibrium decision and pricing functions determine the equilibrium realization for that period. Equilibrium investment and the new technology shocks determine next period’s state. In the next and subsequent periods, this procedure is repeated until time series of the desired length are obtained. The resulting model time series can then be summarized by a suitable set of statistics. In Kydland and Prescott (1982), we built a model economy where all fluctuations could be traced back to technology shocks. We began by extending the neoclassical growth model to include leisure as an argument of the stand-in household’s utility function. Given that more than half of business cycle fluctuations are accounted for by variations in the labor input, introducing this element was crucial. We then calibrated the deterministic version of the model so that its consumption-investment shares, factor income shares, capital/output ratios, leisure/market-time shares and depreciation shares matched the average values for the U.S. economy in the postwar period. We abstracted from public finance considerations and consolidated the public and private sectors. We introduced Frisch’s (1933b) assumption of time to build new productive capital. The construction period considered was four quarters, with new capital becoming productive only upon completion, but with resources being used up throughout the construction period. Given the high volatility of inventory investment, inventory stocks were included as a factor of production. In our computational experiments, using technology shock variance estimated from production function residuals (Prescott, 1986), we found that the model economy’s output variance was 55 percent as large as the corresponding variance for the U.S. economy in the postwar period. Probably the most questionable assumption of this theory is that of homogeneous workers, with the additional implication that all variation in hours occurs in the form of changes in hours per worker. According to aggregate data for the U.S. economy, only about one-third of the quarterly fluctuations in market hours are of this form, while the remaining two-thirds arise from changes in the number of workers (Kydland and Prescott, 1990, Table 1). This observation led Hansen (1985) to introduce the Rogerson (1988) labor indivisibility construct into a business cycle model. In the Hansen world, all fluctuations in hours are in the form of employment variation. To deal with the apparent nonconvexity arising from the assumption of indivisible labor, Hansen makes the problem convex by assuming that the commodity points are contracts in which every agent is paid the same amount whether that agent works or not, and that a lottery randomly chooses who actually works in every period. He finds that with this labor indivisibility, his model economy fluctuates as much as did the U.S. economy. Our view is that with the extreme assumption of fluctuations only in employment, Hansen overestimates the amount of aggregate fluctuations accounted for by Solow residuals in the same way that our equally extreme assumption of fluctuations solely in hours per worker led to an underestimation. In Kydland and Prescott (1991b), the major improvement on the 1982 version of the model economy is that variation is permitted in both the © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 250 82 METHODOLOGY Journal of Economic Perspectives number of workers and the number of hours per worker. The number of hours in which a plant is operated in any given period is endogenous. Because the cost of moving workers in and out of the labor force is not included, a property of the equilibrium is that all of the hours variation is in the form of employment change and none in hours per worker. In that respect, the Kydland and Prescott (1991b) model is identical to Hansen’s (1985) model. Using the economy with no moving costs, technology shocks are estimated to account for about 90 percent of the aggregate output variance. For the economy with moving costs, we calibrated it so that relative variations in hours per worker and the number of workers matched U.S. data. The estimate of the fraction of the cycle accounted for by technology shocks is then reduced to 70 percent. These estimates of the contribution of technology shocks to aggregate fluctuations have been found to be robust in several modifications of the model economy. For instance, Greenwood, Hercowitz and Huffman (1988) permit the utilization rate of capital to vary and affect its depreciation rate and assume all technology change is embodied in new capital. Danthine and Donaldson (1990) introduce an efficiency-wage construct, while Cho and Cooley (1995) permit nominal-wage contracting. Ríos-Rull (1994) uses a model calibrated to life cycle earnings and consumption patterns. Gomme and Greenwood (1995) incorporate heterogeneous agents with recursive preferences and equilibrium risk allocations. In none of these cases is the estimate of the contribution of technology shocks to aggregate fluctuations significantly altered. The computational experiment is also being used to derive the quantitative implications of monetary shocks for business cycle fluctuations if money is used for transaction purposes only. In these experiments, money may be held either because it is required in advance of purchasing cash goods (Lucas and Stokey, 1987; Cooley and Hansen 1989, 1992) or because real cash balances economize on time (Kydland, 1989). Models of this type have been used to evaluate monetary policy. At this stage, we are less confident in these model economies than those designed to evaluate the contribution of technology shocks. There are three related reasons. The first is that, unlike actual economies, these model economies fail to display the sluggishness of the inflation rate’s response to changes in the growth rate of money (Christiano and Eichenbaum, 1992b). The second is that people seem to hold far larger monetary assets than are needed for transaction purposes—in the postwar period, for example, U.S. households’ holdings of M2 have exceeded half of annual GNP—which implies that the transaction rationale for holding money is not well understood. The third reason is that the evaluation of monetary policy appears to be sensitive to the reason people hold these liquid assets. Imrohoroglu (1992) has constructed a model economy in which people vary their holdings of liquid assets as their income varies to smooth their consumption.11 She finds that if a transaction-cost model is calibrated to data generated by her model economy and the calibrated 11 Imrohoroglu and Prescott (1991) introduce a banking technology to intermediate government debt. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE COMPUTATIONAL EXPERIMENT Finn E.Kydland and Edward C.Prescott 251 83 economy is used to estimate the cost of inflation, this estimate would grossly underestimate the true cost of inflation for her model world. This result is surprising and bothersome. Typically, how some feature is introduced is unimportant as long as the aggregate substitution elasticities and quantities match. We currently do not have the tools for computing equilibria of models with both the features of the neoclassical growth model and the idiosyncratic shocks that result in people holding money for precautionary reasons. One may say that stronger theory is needed when it comes to evaluating the contribution of monetary policy shocks to business cycle fluctuations. Summary With the general equilibrium approach, empirical knowledge is organized around preferences and technologies. Given the question and given existing economic theory and measurement, a researcher creates a model economy. This researcher then determines a quantitative answer to the posed question for the model economy. If the theory is strong and the measurements good, we have confidence that the answer for the model economy will be essentially the same as for the actual economy. Of course, sometimes measurement is not very good, and a series of computational experiments reveals that different plausible values of some parameter give very different answers to the posed question. If so, this parameter—which measures some aspect of people’s willingness and ability to substitute—must be more accurately measured before theory can provide an answer in which we have confidence. Or sometimes the theory relative to the question is weak or nonexistent, and the answer depends upon which of the currently competing theories is used to construct the model economy. If so, these competing theories must be subjected to further tests before there is a good basis for choosing among them. At still other times, the computational tools needed to derive the implications of the theory do not exist, so better computational methods or more powerful computers are needed. Earlier in this article, we quoted the Lucas (1980) definition of “theory” as being an explicit set of instructions for building an imitation economy to address certain questions and not a collection of assertions about the behavior of the actual economy. Consequently, statistical hypothesis testing, which is designed to test assertions about actual systems, is not an appropriate tool for testing economic theory. One way to test a theory is to determine whether model economies constructed according to the instructions of that theory mimic certain aspects of reality. Perhaps the ultimate test of a theory is whether its predictions are confirmed—that is, did the actual economy behave as predicted by the model economy, given the policy rule selected? If a theory passes these tests, then it is tested further, often by conducting a computational experiment that includes some feature of reality not previously included in the computational experiments. More often than not, introducing this feature does not change the answers, and currently established theory becomes stronger. Occasionally, however, the new feature turns out to be important, and established theory © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 252 84 METHODOLOGY Journal of Economic Perspectives must then be expanded and improved. In this way, economic science progresses. Given the infeasibility of controlled experiments with national economies, the computational experiment is the tool of quantitative economic theory, whether the primary concern be with theory use or with theory development and testing. We thank Graham Candler, Javier Díaz-Giménez, Tryphon Kollintzas, Jim Schmitz and Nancy Stokey for helpful discussions and the National Science Foundation for financial support. The views expressed herein are those of the authors and not necessarily those of the Federal Reserve Banks of Cleveland or Minneapolis, nor of the Federal Reserve System. References Aiyagari, S.Rao, “On the Contribution of Technology Shocks to Business Cycles,” Quarterly Review, Federal Reserve Bank of Minneapolis, Winter 1994, 18, 22–34. Auerbach, Alan J., and Laurence J.Kotlikoff, Dynamic Fiscal Policy. Cambridge, U.K.: Cambridge University Press, 1987. Backus, David K., Patrick J.Kehoe, and Finn E.Kydland , “Dynamics of the Trade Balance and the Terms of Trade: The J-curve?,” American Economic Review, March 1994, 84, 84–103. Benhabib, Jess, Richard Rogerson, and Randall Wright, “Homework in Macroeconomics: Household Production and Aggregate Fluctuations,” Journal of Political Economy, December 1991, 99, 1166–87. Brown, Drussilla K., Alan V.Deardorff, and Robert M.Stern, “Estimates of a North American Free Trade Agreement,” manuscript, Federal Reserve Bank of Minneapolis, 1994. Chang, Ly-June, “Business Cycles with Distorting Taxes and Disaggregate Capital Markets,” Journal of Economic Dynamics and Control, July/September 1995, 19, 985–1010. Cho, Jong Ok, and Thomas F.Cooley, “The Business Cycle with Nominal Contracts,” Economic Theory, June 1995, 6, 13–34. Christiano, Lawrence J., and Martin Eichenbaum, “Current Real-Business-Cycle Theories and Aggregate Labor-Market Fluctuations,” American Economic Review, June 1992a, 82, 430–50. Christiano, Lawrence J., and Martin Eichenbaum, “Liquidity Effects, Monetary Policy, and the Business Cycle.” Working Paper No. 4129, National Bureau of Economic Research, 1992b. Cooley, Thomas F., and Gary D.Hansen, “The Inflation Tax in a Real Business Cycle Model,” American Economic Review, September 1989, 79, 733–48. Cooley, Thomas F., and Gary D.Hansen, “Tax Distortions in a Neoclassical Monetary Economy,” Journal of Economic Theory, December 1992, 58, 290–316. Danthine, Jean-Pierre, and John B.Donaldson, “Efficiency Wages and the Business Cycle Puzzle,” European Economic Review, November 1990, 34, 1275–1301. Finn, Mary G., “Variance Properties of Solow’s Productivity Residual and their Cyclical Implications,” Journal of Economic Dynamics and Control, July/September 1995, 19, 1249–82. Frisch, Ragnar, “Editorial,” Econometrica, January 1933a, 1, 1–5. Frisch, Ragnar, “Propagation Problems and Impulse Problems in Dynamic Economics.” In Economic Essays in Honour of Gustav Cassel. London: G.Allen & Unwin Ltd., 1933b, pp. 171–205. Ghez, Gilbert R., and Gary S.Becker, The Allocation of Time and Goods over the Life Cycle. New York: Columbia University Press, 1975. Gomme, Paul, and Jeremy Greenwood, “On the Cyclical Allocation of Risk,” Journal of Economic Dynamics and Control, January/ February 1995, 19, 91–124. Gravelle, Jane G., and Laurence J.Kotlikoff, “Corporate Taxation and the Efficiency Gains of the 1986 Tax Reform Act,” Economic Theory, June 1995, 6, 51–82. Greenwood, Jeremy, and Zvi Hercowitz, “The Allocation of Capital and Time Over the Business Cycle,” Journal of Political Economy, December 1991, 99, 1188–214. Greenwood, Jeremy, Zvi Hercowitz, and Gregory W.Huffman, “Investment, Capacity Utilization, and the Real Business Cycle,” American Economic Review, June 1988, 78, 402–18. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors THE COMPUTATIONAL EXPERIMENT 253 The Computational Experiment: An Econometric Tool 85 Gregory, Allan W., and Gregory W.Smith, “Statistical Aspects of Calibration in Macroeconomics.” In Maddala, G.S., C.R.Rao, and H.D.Vinod, eds., Handbook of Statistics. Vol. 2, New York: Elsevier Science, 1993, pp. 703–19. Hansen, Gary D., “Indivisible Labor and the Business Cycle,” Journal of Monetary Economics, November 1985, 16, 309–27. Hansen, Gary D., “Growth and Fluctuations.” In “Three Essays on Labor Indivisibility and the Business Cycle,” Ph.D. dissertation, University of Minnesota, 1986. Harris, Richard, “Applied General Equilibrium Analysis of Small Open Economies with Scale Economies and Imperfect Competition,” American Economic Review, December 1984, 18, 1016–32. Hodrick, Robert J., and Edward C.Prescott, “Post-War U.S. Business Cycles: An Empirical Investigation.” Discussion Paper No. 451, Carnegie-Mellon University, 1980. Hopenhayn, Hugo, and Richard Rogerson, “Job Turnover and Policy Evaluation: A General Equilibrium Analysis,” Journal of Political Economy, October 1993, 101, 915–38. Hornstein, Andreas, “Monopolistic Competition, Increasing Returns to Scale and the Importance of Productivity Change,” Journal of Monetary Economics, June 1993, 31, 299–316. Imrohoroglu, Ayse, “The Welfare Cost of Inflation Under Imperfect Insurance,” Journal of Economic Dynamics and Control, January 1992, 16, 79–91. Imrohoroglu, Ayse, and Edward C.Prescott, “Seigniorage as a Tax: A Quantitative Evaluation, “Journal of Money, Credit and Banking, August 1991, 23, 462–75. Kuhn, Thomas S., The Structure of Scientific Revolutions. Chicago: University of Chicago Press, 1962. Kydland, Finn E., “The Role of Money in a Business Cycle Model.” Discussion Paper No. 23, Institute for Empirical Macroeconomics, Federal Reserve Bank of Minneapolis, 1989. Kydland, Finn E., and Edward C.Prescott, “Time to Build and Aggregate Fluctuations,” Econometrica, November 1982, 50, 1345–70. Kydland, Finn E., and Edward C.Prescott, “Business Cycles: Real Facts and a Monetary Myth,” Quarterly Review, Federal Reserve Bank of Minneapolis, Spring 1990, 14, 3–18. Kydland, Finn E., and Edward C.Prescott, “The Econometrics of the General Equilibrium Approach to Business Cycles,” Scandinavian Journal of Economics, 1991a, 93:2, 161–78. Kydland, Finn E., and Edward C.Prescott, “Hours and Employment Variation in Business Cycle Theory,” Economic Theory, January 1991b, 1, 63–81. Lucas, Robert E., Jr., “Methods and Problems in Business Cycle Theory,” Journal of Money, Credit and Banking, November 1980, 12, 696–715. Lucas, Robert E., Jr., and Nancy L.Stokey, “Money and Interest in a Cash-inAdvance Economy,” Econometrica, May 1987, 55, 491–513. Marschak, Jacob, “Introduction to Econometrics,” hectographed lecture notes, University of Buffalo, 1948. McCallum, Bennett T., “Real Business Cycle Models.” In Barro, Robert J., ed., Modern Business Cycle Theory. Cambridge, Mass.: Harvard University Press, 1989, pp. 16–50. Prescott, Edward C., “Theory Ahead of Business Cycle Measurement,” Quarterly Review, Federal Reserve Bank of Minneapolis, Fall 1986, 10, 9–22; also in Brunner, Karl, and Allan H.Meltzer, eds., Real Business Cycles, Real Exchange Rates, and Actual Policies, Carnegie-Rochester Conference Series on Public Policy. Vol. 25, Amsterdam: North-Holland, 1986, pp. 11–44. Ríos-Rull, José-Víctor, “On the Quantitative Importance of Market Completeness,” Journal of Monetary Economics, December 1994, 34, 462–96. Ríos-Rull, José-Víctor, “Models with Heterogeneous Agents.” In Cooley, T.F., ed., Frontiers of Business Cycle Research. Princeton, N.J.: Princeton University Press, 1995. Rogerson, Richard, “Indivisible Labor, Lotteries and Equilibrium,” Journal of Monetary Economics, January 1988, 21, 3–16. Schneider, Stephen H., “Climate Modeling,” Scientific American, May 1987, 256, 72–80. Shoven, John B., and J.Whalley, “A General Equilibrium Calculation of the Differential Taxation of Income from Capital in the U.S.,” Journal of Public Economics, November 1972, 1, 281–321. Solow, Robert M., Growth Theory: An Exposition. Oxford: Clarendon Press, 1970. Stokey, Nancy L., and Robert E.Lucas, Jr. (with Edward C.Prescott), Recursive Methods in Economic Dynamics. Cambridge, Mass.: Harvard University Press, 1989. Wicksell, Knut, “Krisernas gåta,” Statsøkonomisk Tidsskrift, 1907, 21, 255–70. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 254 CHAPTER 14 Journal of Economic Perspectives—Volume 10, Number 1—Winter 1996—Pages 87–104 The Empirical Foundations of Calibration Lars Peter Hansen and James J.Heckman G eneral equilibrium theory provides the intellectual underpinnings for modern macroeconomics, finance, urban economics, public finance and numerous other fields. However, as a paradigm for organizing and synthesizing economic data, it poses some arduous challenges. A widely accepted empirical counterpart to general equilibrium theory remains to be developed. There are few testable implications of general equilibrium theory for a time series of aggregate quantities and prices. There are a variety of ways to formalize this claim. Sonnenschein (1973) and Mantel (1974) show that excess aggregate demand functions can have “arbitrary shapes” as long as Walras’ Law is satisfied. Similarly, Harrison and Kreps (1979) show that a competitive equilibria can always be constructed to rationalize any arbitrage-free specification of prices. Observational equivalence results are pervasive in economics. There are two responses to this state of affairs. One can view the flexibility of the general equilibrium paradigm as its virtue. Since it is hard to reject, it provides a rich apparatus for interpreting and processing data. 1 Alternatively, general equilibrium theory can be dismissed as being empirically irrelevant because it imposes no testable restrictions on market data. Even if we view the “flexibility” of the general equilibrium paradigm as a virtue, identification of preferences and technology is problematic. Concern 1 Lucas and Sargent (1988) make this point in arguing that early Keynesian critiques of classical economics were misguided by their failure to recognize this flexibility. Lars Peter Hansen is Homer J.Livingston Professor of Economics, and James Heckman is Henry Schultz Distinguished Service Professor of Economics and Director of the Center for Social Program Evaluation at the Irving B.Harris School of Public Policy Studies, all at the University of Chicago, Chicago, Illinois. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors EMPIRICAL FOUNDATIONS OF CALIBRATION 88 255 Journal of Economic Perspectives about the lack of identification of aggregate models has long troubled econometricians (for example, Liu, 1960; Sims, 1980). The tenuousness of identification of many models makes policy analysis and the evaluation of the welfare costs of programs a difficult task and leads to distrust of aggregate models. Different models that “fit the facts” may produce conflicting estimates of welfare costs and dissimilar predictions about the response of the economy to changes in resource constraints. Numerous attempts have been made to circumvent this lack of identification, either by imposing restrictions directly on aggregate preferences and technologies, or by limiting the assumed degree of heterogeneity in preferences and technologies. For instance, the constant elasticity of substitution specification for preferences over consumption in different time periods is one of workhorses of dynamic stochastic equilibrium theory. When asset markets are sufficiently rich, it is known from Gorman (1953) that these preferences can be aggregated into the preferences of a representative consumer (Rubinstein, 1974). Similarly, CobbDouglas aggregate production functions can be obtained from Leontief micro technologies aggregated by a Pareto distribution for micro productivity parameters (Houthakker, 1956). These results give examples of when simple aggregate relations can be deduced from relations underlying the micro behavior of the individual agents, but they do not justify using the constructed aggregate relations to evaluate fully the welfare costs and benefits of a policy.2 Micro data offer one potential avenue for resolving the identification problem, but there is no clear formal statement that demonstrates how access to such data fully resolves it. At an abstract level, Brown and Matzkin (1995) show how to use information on individual endowments to obtain testable implications in exchange economies. As long as individual income from endowments can be decomposed into its component sources, they show that the testability of general equilibrium theory extends to production economies. Additional restrictions and considerable price variation are needed to identify microeconomic preference relations for data sets that pass the Brown-Matzkin test. Current econometric practice in microeconomics is still far from the nonparametric ideal envisioned by Brown and Matzkin (1995). As shown by Gorman (1953), Wilson (1968), Aigner and Simon (1970) and Simon and Aigner (1970), it is only under very special circumstances that a micro parameter such as the intertemporal elasticity of substitution or even a marginal propensity to consume out of income can be “plugged into” a representative consumer model to produce an empirically concordant aggregate model. As illustrated by Houthakker’s (1956) result, microeconomic technologies can look quite different from their aggregate counterparts. In practice, microeconomic elasticities are often estimated by reverting to a partial equilibrium econometric model. Cross-market price 2 Gorman’s (1953) results provide a partial justification for using aggregate preferences to compare alternative aggregate paths of the economy. Even if one aggregate consumptioninvestment profile is preferred to another via this aggregate preference ordering, to convert this into a Pareto ranking for the original heterogeneous agent economy requires computing individual allocations for the path—a daunting task. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 256 METHODOLOGY Lars Peter Hansen and James J.Heckman 89 elasticities are either assumed to be zero or are collapsed into constant terms or time dummies as a matter of convenience. General equilibrium, multimarket price variation is typically ignored in most microeconomic studies. Battle lines are drawn over the issue of whether the microeconometric simplifications commonly employed are quantitatively important in evaluating social welfare and assessing policy reforms. Shoven and Whalley (1972, 1992) attacked Harberger’s use of partial equilibrium analysis in assessing the effects of taxes on outputs and welfare. Armed with Scarf’s algorithm (Scarf and Hansen, 1973), they computed fundamentally larger welfare losses from taxation using a general equilibrium framework than Harberger computed using partial equilibrium analysis. However, these and other applications of general equilibrium theory are often greeted with skepticism by applied economists who claim that the computations rest on weak empirical foundations. The results of many simulation experiments are held to be fundamentally implausible because the empirical foundations of the exercises are not secure. Kydland and Prescott are to be praised for taking the general equilibrium analysis of Shoven and Whalley one step further by using stochastic general equilibrium as a framework for understanding macroeconomics.3 Their vision is bold and imaginative, and their research program has produced many creative analyses. In implementing the real business cycle program, researchers deliberately choose to use simple stylized models both to minimize the number of parameters to be “calibrated” and to facilitate computations.4 This decision forces them to embrace a rather different notion of “testability” than used by the other general equilibrium theorists, such as Sonnenschein, Mantel, Brown and Matzkin. At the same time, the real business cycle community dismisses conventional econometric testing of parametric models as being irrelevant for their purposes. While Kydland and Prescott advocate the use of “well-tested theories” in their essay, they never move beyond this slogan, and they do not justify their claim of fulfilling this criterion in their own research. “Well tested” must mean more than “familiar” or “widely accepted” or “agreed on by convention,” if it is to mean anything. Their suggestion that we “calibrate the model” is similarly vague. On one hand, it is hard to fault their advocacy of tightly parameterized models, because such models are convenient to analyze and easy to understand. Aggregate growth coupled with uncertainty makes nonparametric identification of preferences and technology extremely difficult, if not impossible. Separability and homogeneity restrictions on preferences and technologies have considerable appeal as identifying assumptions. On the other hand, Kydland and Prescott never provide a coherent framework for 3 The earlier work by Lucas and Prescott (1971) took an initial step in this direction by providing a dynamic stochastic equilibrium framework for evaluating empirical models of investment. 4 The term “real business cycle” originates from an emphasis on technology shocks as a source of business cycle fluctuations. Thus, real, as opposed to nominal, variables drive the process. In some of the recent work, both real and nominal shocks are used in the models. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors EMPIRICAL FOUNDATIONS OF CALIBRATION 90 257 Journal of Economic Perspectives extracting parameters from microeconomic data. The same charge of having a weak empirical foundation that plagued the application of deterministic general equilibrium theory can be lodged against the real business cycle research program. Such models are often elegant, and the discussions produced from using them are frequently stimulating and provocative, but their empirical foundations are not secure. What credibility should we attach to numbers produced from their “computational experiments,” and why should we use their “calibrated models” as a basis for serious quantitative policy evaluation? The essay by Kydland and Prescott begs these fundamental questions. The remainder of our essay is organized as follows. We begin by discussing simulation as a method for understanding models. This method is old, and the problems in using it recur in current applications. We then argue that model calibration and verification can be fruitfully posed as econometric estimation and testing problems. In particular, we delineate the gains from using an explicit econometric framework. Following this discussion, we investigate current calibration practice with an eye toward suggesting improvements that will make the outputs of computational experiments more credible. The deliberately limited use of available information in such computational experiments runs the danger of making many economic models with very different welfare implications compatible with the evidence. We suggest that Kydland and Prescott’s account of the availability and value of micro estimates for macro models is dramatically overstated. There is no filing cabinet full of robust micro estimates ready to use in calibrating dynamic stochastic general equilibrium models. We outline an alternative paradigm that, while continuing to stress the synergy between microeconometrics and macro simulation, will provide more credible inputs into the computational experiments and more accurate assessments of the quality of the outputs. Simulation as a Method for Understanding Models In a simple linear regression model, the effect of an independent variable on the dependent variable is measured by its associated regression coefficient. In the dynamic nonlinear models used in the Kydland-Prescott real business cycle research program, this is no longer true. The dynamic nature of such models means that the dependent variable is generated in part from its own past values. Characterizing the dynamic mechanisms by which exogenous impulses are transmitted into endogenous time series is important to understanding how these models induce fluctuations in economic aggregates. Although there is a reliance on linearization techniques in much of the current literature, for large impulses or shocks, the nonlinear nature of such models is potentially important. To capture the richness of a model, the analyst must examine various complicated features of it. One way to do this is to simulate the model at a variety of levels of the forcing processes, impulses and parameters. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 258 METHODOLOGY The Empirical Foundations of Calibration 91 The idea of simulating a complex model to understand its properties is not a new principle in macroeconomics. Tinbergen’s (1939) simulation of his League of Nations model and the influential simulations of Klein and Goldberger (1955) and Goldberger (1959) are but three of a legion of simulation exercises performed by previous generations of economists.5 Fair (1994) and Taylor (1993) are recent examples of important studies that rely on simulation to elucidate the properties of estimated models. However, the quality of any simulation is only as good as the input on which it is based. The controversial part of the real business cycle simulation program is the method by which the input parameters are chosen. Pioneers of simulation and of economic dynamics like Tinbergen (1939) and Frisch (1933) often guessed at the parameters they used in their models, either because the data needed to identify the parameters were not available, or because the econometric methods were not yet developed to fully exploit the available data (Morgan, 1990). At issue is whether the state of the art for picking the parameters to be used for simulations has improved since their time. Calibration versus Estimation A novel feature of the real business cycle research program is its endorsement of “calibration” as an alternative to “estimation.” However, the distinction drawn between calibrating and estimating the parameters of a model is artificial at best.6 Moreover, the justification for what is called “calibration” is vague and confusing. In a profession that is already too segmented, the construction of such artificial distinctions is counterproductive. It can only close off a potentially valuable dialogue between real business cycle research and other research in modern econometrics. Since the Kydland-Prescott essay is vague about the operating principles of calibration, we turn elsewhere for specificity. For instance, in a recent description of the use of numerical models in the earth sciences, Oreskes, Shrader-Frechette and Belitz (1994, pp. 642, 643) describe calibration as follows: In earth sciences, the modeler is commonly faced with the inverse problem: The distribution of the dependent variable (for example, the hydraulic head) is the most well known aspect of the system; the distribution of the independent variable is the least well known. The process of tuning the model—that is, the manipulation of the independent variables to obtain a match between the observed and simulated distribution or distributions of a dependent variable or variables—is known as calibration. 5 Simulation is also widely used in physical science. For example, it is customary in the studies of fractal dynamics to simulate models in order to gain understanding of the properties of models with various parameter configurations (Peitgen and Richter, 1986). 6 As best we can tell from their essay, Kydland and Prescott want to preserve the term “estimation” to apply to the outputs of their computational experiments. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors EMPIRICAL FOUNDATIONS OF CALIBRATION 92 259 Journal of Economic Perspectives Some hydrologists have suggested a two-step calibration scheme in which the available dependent data set is divided into two parts. In the first step, the independent parameters of the model are adjusted to reproduce the first part of the data. Then in the second step the model is run and the results are compared with the second part of the data. In this scheme, the first step is labeled “calibration” and the second step is labeled “verification.” This appears to be an accurate description of the general features of the “calibration” method advocated by Kydland and Prescott. For them, data for the first step come from micro observations and from secular growth observations (see also Prescott, 1986a). Correlations over time and across variables are to be used in the second step of verification. Econometricians refer to the first stage as estimation and the second stage as testing. As a consequence, the two-stage procedure described by Oreskes, Shrader-Frechette and Belitz (1994) has a straightforward econometric counterpart.7 From this perspective, the Kydland-Prescott objection to mainstream econometrics is simply a complaint about the use of certain loss functions for describing the fit of a model to the data or for producing parameter estimates. Their objection does not rule out econometric estimation based on other loss functions. Econometric estimation metrics like least squares, weighted least squares or more general method-of-moments metrics are traditional measures of fit. Difference among these methods lie in how they weight various features of the data; for example, one method might give a great deal of weight to distant outliers or to certain variables, causing them to pull estimated trend lines in their direction; another might give less weight to such outliers or variables. Each method of estimation can be justified by describing the particular loss function that summarizes the weights put on deviations of a model’s predictions from the data. There is nothing sacred about the traditional loss functions in econometrics associated with standard methods, like ordinary least squares. Although traditional approaches do have rigorous justifications, a variety of alternative loss functions could be explored that weight particular features of a model more than others. For example, one could estimate with a loss function that rewarded models that are more successful in predicting turning points. Alternatively, particular time series frequencies could be deemphasized in adopting an estimation criterion because misspecification of a model is likely to contaminate some frequencies more than others (Dunsmuir and Hannan, 1978; Hansen and Sargent, 1993; Sims, 1993). 7 See Christiano and Eichenbaum (1992) for one possible econometric implementation of this two-step approach. They use a generalized method of moments formulation (for example, Hansen, 1982) in which parameters are estimated by a first stage, exactly identified set of moment relations, and the model is tested by looking at another set of moment restrictions. Not surprisingly, to achieve identification of the underlying set of parameters, they are compelled to include more than just secular growth relations in the first-stage estimation, apparently violating one of the canons of current calibration practice. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 260 METHODOLOGY Lars Peter Hansen and James J.Heckman 93 The real business cycle practitioners adopt implicit loss functions. In looking at economic aggregates, their implicit loss functions appear to focus on the model predictions for long-run means, to the exclusion of other features of the data, when selecting parameter estimates. It is unfortunate that we are forced to guess about the rationale for the loss functions implicit in their research. There is little emphasis on assessing the quality of the resulting calibration. Formalizing the criteria for calibration and verification via loss functions makes the principle by which a particular model is chosen easier to understand. A clear statement would lead to more fruitful and focused conversations about the sources and reliability of estimated parameters. As Oreskes, Shrader-Frechette and Belitz (1994) emphasize, the distinction between calibration and verification is often contrived. In many circumstances the verification step should really be considered part of the “calibration” step. The absence of a sharp distinction between these two stages is consistent with the difficulty of obtaining testable implications from the general equilibrium paradigm. Model testing serves as a barometer for measuring whether a given parametric structure captures the essential features of the data. When cleverly executed, it can pinpoint defective features of a model. Applied statistical decision theory and conventional statistical practice provide a formalism for conducting this endeavor. While this theory can be criticized for its rigidity or its naiveté, it seems premature to scrap it altogether without putting in place some other clearly stated criterion for picking the parameters of a model and assessing the quality of that selection. The rational agents in a model of the Kydland-Prescott type rely explicitly on loss functions. After all, their rational decision making is based on the application of statistical decision theory, and part of the Kydland-Prescott line of research is to welcome the application of this theory to modern macroeconomics. But the idea of a loss function is also a central concept in statistical decision theory (LeCam and Yang, 1990). The rational agents in real business cycle models use this theory and, as a consequence, are assumed to process information in a highly structured way. Why should the producers of estimates for the real business cycle models act differently? Although Kydland and Prescott are not precise in this essay in stating how calibration should be done in practice, there is much more specificity in Prescott (1986a, p. 14), who writes: “The key parameters of the growth model are the intertemporal and intratemporal elasticities of substitution. As Lucas (1980, p. 712) emphasizes, ‘On these parameters, we have a wealth of inexpensively available data from census and cohort information, from panel data describing market conditions and so forth.’” It is instructive to compare Prescott’s optimistic discussion of the ease of using micro data to inform calibration with the candid and informative discussion of the same issue by Shoven and Whalley (1992, p. 105), who pioneered the application of calibration methods in general equilibrium analysis. They write: Typically, calibration involves only one year’s data or a single observation represented as an average over a number of years. Because of the reliance on a single observation, benchmark data typically does not identify a unique © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors EMPIRICAL FOUNDATIONS OF CALIBRATION 94 261 Journal of Economic Perspectives set of values for the parameters in any model. Particular values for the relevant elasticities are usually required, and are specified on the basis of other research. These serve, along with the equilibrium observation, to uniquely identify the other parameters of the model. This typically places major reliance on literature surveys of elasticities; as many modelers have observed in discussing their own work, it is surprising how sparse (and sometimes contradictory) the literature is on some key elasticity values. And, although this procedure might sound straightforward, it is often exceedingly difficult because each study is different from every other. What is noteworthy about this quotation is that the authors are describing a deterministic general equilibrium model based on traditional models of factor demand, sectoral output, product supply, labor supply and demand for final products, which have been the focus of numerous micro empirical studies. There have been many fewer micro empirical studies of the sectoral components of the stochastic general equilibrium models used in real business cycle theory. If there are few well-tested models that Shoven and Whalley can pull off the shelf, is it plausible that the shelf is unusually rich in models estimated assuming the relevant economic agents are operating in the more general economic environments considered in real business cycle theory? Shoven and Whalley (1992, p. 106) come close to acknowledging the fundamental underidentification of general equilibrium systems from time series data when they write: [I]n some applied models many thousands of parameters are involved, and to estimate simultaneously all of the model parameters using timeseries methods would require either unrealistically large numbers of observations or overly severe identifying restrictions.… Thus far, these problems have largely excluded complete econometric estimation of general equilibrium systems in applied work. Current real business cycle models often require many fewer parameters to be calibrated, because they are highly aggregated. However, the extraction of the required elasticities from microeconometric analyses is more problematic, because the implicit economic environments invoked to justify microeconometric estimation procedures seldom match the dynamic stochastic single-agent models for which the micro estimates act as inputs. Microeconomic studies rarely estimate models that can be directly applied to the aggregates used in real business cycle theory. Moreover, as the specification of the real business cycle models become richer, they will inevitably have to face up to the same concerns that plague Shoven and Whalley.8 8 This problem has already surfaced in the work of Benhabib, Rogerson and Wright (1991). They try to identify the parameters of a household production function for the services from durable goods using Panel Survey of Income Dynamics data, but without data on one of the inputs (the stock of durable goods), poor data on the other input (time spent by the household required to make durable goods productive) and no data on the output. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 262 METHODOLOGY The Empirical Foundations of Calibration 95 The Real Business Cycle Empirical Method In Practice Kydland and Prescott, along with other real business cycle practitioners, endorse the use of time series averages—but not correlations—in calibrating models. In their proposed paradigm for empirical research, correlations are to be saved and used to test models, but are not to be used as a source of information about parameter values. It has become commonplace in the real business cycle research program to match the steady-state implications of models to time series averages. To an outsider, this looks remarkably like a way of doing estimation without accounting for sampling error in the sample means. In fact, the real business cycle “calibration” estimator of the Cobb-Douglas share parameter is a classical econometric estimator due to Klein (Klein, 1953; Nerlove, 1965). The only difference is that the Klein estimator usually is presented with a standard error. Why is it acceptable to use sample means as a valid source of information about model parameters and not sample autocorrelations and cross correlations? Many interesting parameters cannot be identified from population means alone. Although the real business cycle literature provides no good reason for not using other sample moments, some reasons could be adduced. For example, one traditional argument for using sample means is that they are robust to measurement error in a way that sample variances and covariances are not as long as the errors have mean zero. Another possible rationale is that steady-state relations are sometimes robust with respect to alternative specifications of the short-run dynamics of a model. In these cases, a calibration fit to sample means will be consistent with a class of models that differ in their implications for short-run dynamics. However, the other side of this coin is that long-term means identify the short-run dynamics of a model only in very special circumstances. Moreover, as pointed out by Sargent (1989), even with measurement error, time series correlations and cross correlations can still provide more information about a model than is conveyed in sample means. Since the models considered by Kydland and Prescott are stochastic, it is not in general possible to calibrate all of the parameters of a model solely from the means of macro time series. Computational experiments make assumptions about the correlation among the stochastic inputs to the model. Information about shocks, such as their variances and serial correlations, are needed to conduct the computational experiments. In a related vein, macroeconomic correlations contain potentially valuable information about the mechanism through which shocks are transmitted to macro time series. For models with richer dynamics, including their original “time-to-build” model, Kydland and Prescott (1982) envision fully calibrating the transmission mechanisms from micro evidence; but they provide no defense for avoiding the use of macro correlations in that task. Recently, Cogley and Nason (1995) have criticized models in the literature spawned by Kydland and Prescott for failing to generate business cycle dynamics (see also Christiano, 1988; Watson, 1993; Cochrane, 1994). Since matching the full set of dynamics of the model to the dynamics in the data is not an essential part of calibration methodology, these models survive the © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors EMPIRICAL FOUNDATIONS OF CALIBRATION 96 263 Journal of Economic Perspectives weak standards for verification imposed by the calibrators. A much more disciplined and systematic exploration of the intertemporal and cross correlations, in a manner now routine in time series econometrics, would have shifted the focus from the empirical successes to the empirical challenges. We agree with Oreskes, Shrader-Frechette and Belitz (1994) that the distinction between calibration and verification is commonly blurred in practice. In the case of real business cycle research, such blurring is likely to be all the more prevalent as the models are redesigned to incorporate richer transient dynamics and additional sources of uncertainty. As Kydland and Prescott emphasize, one of the most important questions for macroeconometrics is the quantitative importance of alternative sources of business cycle fluctuations. This classical problem has not yet been definitively answered (Morgan, 1990, pt. I). Using intuition from factor analysis, it is impossible to answer this question from a single time series. From two time series, one can isolate a single common factor. (Intuitively, two random variables can always be decomposed into a common component and two uncorrelated components.) Only using multiple time series is it possible to sort out multiple sources of business cycle shocks. The current emphasis in the literature on using only a few “key correlations” to check a model’s implications makes single-factor explanations more likely to emerge from real business cycle analyses.9 The idiosyncratic way Kydland and Prescott quantify the importance of technology shocks unfortunately makes it difficult to compare their answers to those obtained from the “innovation accounting” methods advocated by Sims (1980) and used extensively in empirical macroeconomics or to those obtained using the dynamic factor models of Geweke (1977) and Sargent and Sims (1977). Kydland and Prescott’s answer to the central question of the importance of technology shocks would be much more credible if it were reinforced by other empirical methodologies. A contrast with John Taylor’s approach to investigating the properties of models is instructive. Taylor’s research program includes the use of computational experiments. It is well summarized in his recent book (Taylor, 1993). Like Kydland and Prescott, Taylor relies on fully specified dynamic models and imposes rational expectations when computing stochastic equilibria. However, in fitting linear models he uses all of the information on first and second moments available in the macro data when it is computationally possible to do so. The econometric methods used in parameter estimation are precisely described. Multiple sources of business cycle shocks are admitted into the model at the outset, and rigorous empirical testing of models appears throughout his analyses.10 9 In private correspondence, John Taylor has amplified this point: “I have found that the omission of aggregate price or inflation data in the Kydland-Prescott second moment exercise creates an artificial barrier between real business cycle models and monetary models. To me, the Granger causality from inflation to output and vice versa are key facts to be explained. But Kydland and Prescott have ignored these facts because they do not fit into their models.” 10 Fair (1994) presents an alternative systematic approach to estimation and simulation, but unlike Taylor, he does not impose rational expectations assumptions. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 264 METHODOLOGY Lars Peter Hansen and James J.Heckman 97 If the Kydland and Prescott real business cycle research program is to achieve empirical credibility, it will have to provide a much more comprehensive assessment of the successes and failures of its models. To convince a wide audience of “outsiders,” the proclaimed successes in real business cycle calibration should not be intertwined with an idiosyncratic and poorly justified way of evaluating models. We sympathize with Fair (1992, p. 141), who writes: Is the RBC [real business cycle] approach a good way of testing models? At first glance it might seem so, since computed paths are being compared to actual paths. But the paths are being compared in a very limited way in contrast to the way that the Cowles Commission approach would compare them. Take the simple RMSE [root mean square error11] procedure. This procedure would compute a prediction error for a given variable for each period and then calculate the RMSE from another structural model or from an autoregressive or vector autoregressive model. I have never seen this type of comparison done for a RBC model. How would, say, the currently best-fitting RBC model compare to a simple first-order autoregressive equation for real GNP in terms of the RMSE criterion? My guess is very poorly. Having the computed path mimic the actual path for a few selected moments is a far cry from beating even a first-order autoregressive equation (let alone a structural model) in terms of fitting the observations well according to the RMSE criterion. The disturbing feature of the RBC literature is there seems to be no interest in computing RMSEs and the like. People generally seem to realize that the RBC models do not fit well in this sense, but they proceed anyway. Specification Uncertainty Underlies the Estimates One of the most appealing features of a research program that builds dynamic macroeconomic models on microeconomic foundations is that it opens the door to the use of micro empirical evidence to pin down macro parameter values. Kydland and Prescott and the entire real business cycle community pay only lip service to the incompatibility between the macroeconomic model used in their computational experiments and the microeconometric models used to secure the simulation parameters. It can be very misleading to plug microeconometric parameter estimates into a macroeconomic model when the economic environments for the two models are fundamentally different. In fact, many of the micro studies that the “calibrators” draw upon do not estimate the parameters required by the models being simulated. 11 RMSE is the square root of the mean of the squared discrepancies between predicted and actual outcomes. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors EMPIRICAL FOUNDATIONS OF CALIBRATION 98 265 Journal of Economic Perspectives This creates specification uncertainty (Learner, 1978). To adequately represent this uncertainty, it is necessary to incorporate the uncertainty about model parameters directly into the outputs of simulations. Standard errors analogous to those presented by Christiano and Eichenbaum (1992) and Eichenbaum (1991) are a useful first step, but do not convey the full picture of model uncertainty. What is required is a sensitivity analysis to see how the simulation outputs are affected by different choices of simulation parameters. Trostel (1993) makes effective use of such a methodology. Consider using the estimates of intertemporal labor supply produced by Ghez and Becker (1975) for simulation purposes.12 Ghez and Becker (1975) estimate the intertemporal substitution of leisure time assuming perfect credit markets, no restrictions on trade in the labor market and no fixed costs of work. This study is important, but like all empirical work in economics, the precise estimates are enveloped by some uncertainty. Moreover, different estimation schemes are required to secure this parameter if there is uninsurable uncertainty in the environment (MaCurdy, 1978). Even looking only at estimates of the intertemporal substitution of leisure based on models that assume that workers can perfectly insure, the point estimates reported in the literature are very imprecisely determined (MaCurdy, 1981; Altonji, 1986). Further, it is not clear how the estimates should be modified to be compatible with the other economic environments including settings that allow for uninsurable uncertainty, transactions costs and restrictions on trades in the market. Current practices in the field of calibration and simulation do not report either estimation error and/or model-specification error. Nor is it a standard feature of real business cycle practice to present formal analyses that explore how sensitive the simulations are to different parameter values. Precise numerical outputs are reported, but with no sense of the confidence that can be placed in the estimates. This produces a false sense of precision. Observationally Equivalent Models Offer Different Predictions about Policy Interventions While putting on empirical “blinders” permits a particular line of research to proceed, looking at too narrow of a range of data makes identification problems more severe. A disturbing feature of current practice in the real business cycle 12 Kydland and Prescott cite Ghez and Becker (1975) as a prime example of the value of microeconomic empirical work. However, their citation misses two central aspects of that work. First, Ghez and Becker (1975) use synthetic cohort data, not panel data as stated by Kydland and Prescott. Second, the interpretation of the Ghez-Becker estimates as structural parameters is predicated on a list of identifying assumptions. These assumptions coupled with the resulting estimates are the most important part of their investigation, not their observation that people sleep eight hours a day, which is what Kydland and Prescott emphasize. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 266 METHODOLOGY The Empirical Foundations of Calibration 99 literature is that models with the same inputs can produce fundamentally different computed values of welfare loss and quantitative assessments of alternative economic policies. Consider the following developments in the field of empirical finance. A frequently noted anomaly is that the observed differences in returns between stocks and bonds are too large to be consistent with models of the preferences commonly used in real business cycle analysis (Hansen and Singleton, 1983; Mehra and Prescott, 1985; Cochrane and Hansen, 1992; Kocherlakota, 1996). One response to these asset-pricing anomalies has been the modification to preferences developed by Epstein and Zin (1989), which breaks the tight link between intertemporal substitution and risk aversion that was maintained in the preceding literature. A parallel advance has been the introduction of intertemporal complementarities such as habit persistence in the preference orderings of consumers (Constantinides, 1990). Hansen, Sargent and Tallarini (1995) find that models with Epstein-Zin type preferences and models without this form of risk sensitivity explain the same quantity data but have fundamentally different implications for the market price of risk (the slope of the mean-standard deviation frontier for asset returns). 13 These “observationally equivalent” preference specifications produce very different estimates of the welfare losses associated with hypothetical policy interventions. The decision by other researchers such as Epstein and Zin to look more broadly at available data and to emphasize model defects instead of successes provoked quantitatively important advances in economic theory. Another competing explanation for the equity premium puzzle is the presence of incomplete markets and transactions costs in asset markets. This explanation is consistent with Prescott’s (1986b, p. 29) earlier argument for ignoring asset market data in real business cycle calibrations: “That the representative agent model is poorly designed to predict differences in borrowing and lending rates…does not imply that this model is not well suited for other purposes—for predicting the consequences of technology shocks for fluctuations in business cycle frequencies, for example.” Heaton and Lucas (1995) quantify the magnitude of transaction costs needed to address the equity-premium puzzle (see also Aiyagari and Gertler, 1991). Prescott may be correct that such models will not help to match “key” correlations in economic aggregates, but this requires documentation. Even if there is robustness of the form hoped for by Prescott, the presence of transactions costs of the magnitude suggested by Heaton and Lucas (1995) are likely to alter the welfare comparisons across different policy experiments in a quantitatively important way.14 This is so because transactions costs prevent heterogeneous consumers from equating marginal rates of substitution and put a wedge between marginal rates of substitution and marginal rates of transformation. 13 Recent work by Campbell and Cochrane (1995) and Boldrin, Christiano and Fisher (1995) suggests a similar conclusion for models with strong intertemporal complementarities. 14 This sensitivity actually occurs in the “bothersome” experiment of Gmrohorgu lu (1992) mentioned by Kydland and Prescott. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors EMPIRICAL FOUNDATIONS OF CALIBRATION 100 267 Journal of Economic Perspectives A More Constructive Research Program The idea of using micro data to enrich the information in macro time series dates back at least to the writings of Tobin (1950). A careful reading of the literature that accompanied his suggestion reveals that his idea was inherently controversial, especially if the micro information is based on cross-section data, and if the behavioral equations are dynamic (Aigner and Simon, 1970; Simon and Aigner, 1970). This issue was revived in the late 1970s and early 1980s when numerous economists attempted to estimate micro labor supply equations to test the Lucas and Rapping (1969) intertemporal substitution hypothesis. The hypothesis rests critically on consumer responses to expected real discounted future wage movements relative to current wages. By providing well-focused economic questions, the Lucas-Rapping model advanced the development of empirical microeconomics by challenging economists to supply answers. Numerous micro studies of labor supply were conducted with an eye toward confirming or discontinuing their hypothesis (Altonji and Ashenfelter, 1980; MaCurdy, 1981; Ham, 1982; Altonji, 1986). However, these studies reveal that even with large micro samples, it is not possible to estimate the parameter of interest precisely. Measurement error in micro data and selection problems often limit the value of the information in the micro data. Macro time series or aggregated cross sections can sometimes solve selection problems that are intractable in micro data (Heckman and Robb, 1985, pp. 168–169, 210–213). Different micro survey instruments produce fundamentally different descriptions of the same phenomena (Smith, 1995). Micro data are no panacea. Moreover, the recent movement in empirical microeconomics away from economic models to “simple descriptive” estimation schemes has reduced the supply of new structural parameters. It is simply not true that there is a large shelf of micro estimates already constructed for different economic environments that can be plugged without modification into a new macro model. In many cases, estimators that are valid in one economic environment are not well suited for another. Given the lessthan-idyllic state of affairs, it seems foolish to look to micro data as the primary source for many macro parameters required to do simulation analysis. Many crucial economic parameters—for example, the effect of product inputs on industry supply—can only be determined by looking at relationships among aggregates. Like it or not, time series evidence remains essential in determining many fundamentally aggregative parameters. A more productive research program would provide clearly formulated theories that will stimulate more focused microeconomic empirical research. Much recent micro research is atheoretical in character and does not link up well with macro general equilibrium theory. For example, with rare exceptions, micro studies treat aggregate shocks as nuisance parameters to be eliminated by some trend or dummy variable procedure.15 A redirection of micro empirical work toward providing input into well-defined general 15 For an exception see Heckman and Sedlacek (1985), who show how cross-section time dummies can be used to estimate the time series of unobserved skill prices in a market model of self-selection. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 268 METHODOLOGY Lars Peter Hansen and James J.Heckman 101 equilibrium models would move discussions of micro evidence beyond discussions of whether wage or price effects exist, to the intellectually more important questions of what the micro estimates mean and how they can be used to illuminate well-posed economic questions. “Calibrators” could make a constructive contribution to empirical economics by suggesting a more symbiotic relationship between the macro general equilibrium model as a synthesizing device and motivating vehicle and the micro evidence as a source of robust parameter values. Recently there has been considerable interest in heterogeneous agent models in the real business cycle literature; Ríos-Rull (1995) offers a nice summary. To us, one of the primary reasons for pushing this line of research is to narrow the range of specification errors in calibrating with microeconomic data. Microeconometric estimates routinely incorporate heterogeneity that is often abstracted from the specification of dynamic, stochastic general equilibrium models. It is remarkable to us that so little emphasis has been given to the transition from micro to macro in the real business cycle literature, given that understanding the distribution of heterogeneity is central to making this transition (Stoker, 1993). The Kydland and Prescott program is an intellectually exciting one. To date, however, the computations produced from it have only illustrated some of the qualitative properties of some dynamic stochastic models and demonstrated the possibility of executing an array of interesting calculations. The real business cycle modeling effort would be more beneficial if it shifted its focus to micro predictions and in this way helped to stimulate research on empirical models that would verify or contradict the macro models. We envision a symbiotic relationship between calibrators and empirical economists in which calibration methods like those used by Frisch, Tinbergen, and Kydland and Prescott stimulate the production of more convincing micro empirical estimates by showing which gaps in our knowledge of micro phenomenon matter and which gaps do not. Calibration should only be the starting point of an empirical analysis of general equilibrium models. In the absence of firmly established estimates of key parameters, sensitivity analyses should be routine in real business cycle simulations. Properly used and qualified simulation methods can be an important source of information and an important stimulus to high-quality empirical economic research. The research program we advocate is not an easy one. However, it will be an informative one. It will motivate micro empirical researchers to focus on economically interesting questions; it will secure the foundations of empirical general equilibrium theory; and, properly executed, it will demonstrate both the gaps and strengths of our knowledge on major issues of public policy. We thank Jennifer Boobar, John Cochrane, Marty Eichenbaum, Ray Fair, Chris Flinn, John Heaton, Bob Lucas, Tom Sargent, Jeff Smith, Nancy Stokey, John Taylor and Grace Tsiang for their valuable comments on this draft. Hansen’s research is supported in part by NSF SBR-94095–01; Heckman is supported by NSF 93–048–0211 and grants from the Russell Sage Foundation and the American Bar Foundation. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors EMPIRICAL FOUNDATIONS OF CALIBRATION 269 102 Journal of Economic Perspectives References Aigner, Dennis, and Julian L.Simon, “A Specification BIAS Interpretation of CrossSection vs. Time-Series Parameter Estimates,” Western Economic Journal, June 1970, 2 2, 144–61. Aiyagari, S.Rao, and Mark Gertler, “Asset Returns with Transitions Costs and Uninsured Individual Risk,” Journal of Monetary Economics, June 1991, 27, 311–31. Altonji, Joseph G., “Intertemporal Substitution in Labor Supply: Evidence from Micro Data,” Journal of Political Economy, June 1986, 94:3, Part 2, S176–S215. Altonji, Joseph G., and Orley Ashenfelter, “Wage Movements and the Labour Market Equilibrium Hypothesis,” Economica, August 1980, 49, 783–824. Benhabib, J., R.Rogerson, and R.Wright, “Homework in Macroeconomics: Household Production and Aggregate Fluctuations,” Journal of Political Economy, December 1991, 99, 1166–87. Boldrin, Michele, Lawrence J.Christiano, and J.D.M. Fisher, “Asset Pricing Lessons for Modeling Business Cycles,” manuscript, August 1995. Brown, D.J., and R.L.Matzkin, “Testable Restrictions on the Equilibrium Manifold,” manuscript, April 1995. Campbell, John, and John H.Cochrane, “By Force of Habit: A Consumption-Based Explanation of Aggregate Stock Market Behavior.” National Bureau of Economic Research Working Paper No. 4995, 1995. Christiano, Lawrence J., “Why Does Inventory Investment Fluctuate So Much?,” Journal of Monetary Economics, March/ May 1988, 21, 247–80. Christiano, Lawrence J., and Martin Eichenbaum, “Current Real Business Cycle Theories and Aggregate Labor Market Fluctuations,” American Economic Review, June 1992, 82, 430–50. Cochrane, John H., “Shocks,” CarnegieRochester Conference. Series on Public Policy, Spring 1994, 41, 295–364. Cochrane, John H., and Lars Peter Hansen, “Asset Pricing Explorations for Macroeconomics.” In Blanchard, O.J., and S.Fischer, eds., NBER Macroeconomics Annual 1992. Cambridge, Mass.: Massachusetts Institute of Technology Press, 1992, pp. 115– 69. Cogley, T., and J.M.Nason, “Effects of the Hodrick-Prescott Filter on Trend and Difference Stationary Time Series Implications for Business Cycle Research,” Journal of Economic Dynamics and Control, January/February 1995, 19, 253–78. Constantinides, George M., “Habit Formation: A Resolution of the Equity Premium Puzzle,” Journal of Political Economy, June 1990, 98, 519–43. Dunsmuir, W., and E.J.Hannan, “Vector Linear Time Series Models,” Advances in Applied Probability, 1978, 8:2, 339–64. Eichenbaum, M., “Real Business Theory: Wisdom or Whimsy,” Journal of Economic Dynamics and Control, October 1991, 15, 607–26. Epstein, Larry G., and Stanley E.Zin, “Substitution, Risk Aversion, and the Temporal Behavior of Consumption and Asset Returns: A Theoretical Framework,” Econometrica, July 1989, 57, 937–69. Fair, Ray, “The Cowles Commission Approach, Real Business Cycle Theories, and New-Keynesian Economics.” In Belongia, M.T., and M.R.Garfinkel, eds., The Business Cycle: Theories and Evidence. Boston: Kluwer Academic, 1992, pp. 133–47. Fair, Ray, Testing Macroeconomic Models. Cambridge: Harvard University Press, 1994. Frisch, Ragnar, “Propagation Problems and Impulse Problems in Dynamic Economics.” In Economic Essays in Honor of Gustav Cassel. London: Allen and Unwin, 1933, pp. 171– 205. Geweke, John, “The Dynamic Factor Analysis of Economic Time Series Models.” In Aigner, D.J., and A.S.Goldberger, eds., Latent Variables in Socio-Economic Models. Amsterdam: North-Holland, 1977, pp. 365– 83. Ghez, G., and Gary Becker, The Allocation of Time and Goods Over the Life Cycle. New York: National Bureau of Economic Research, 1975. Goldberger, Arthur, Impact Multipliers and Dynamic Properties of the Klein-Golberger Model. Amsterdam: North-Holland, 1959. German, William M., “Community Preference Fields,” Econometrica, January 1953, 21, 63–80. Ham, John C., “Estimation of a Labour Supply Model with Censoring Due to Unemployment and Underemployment,” Review of Economic Studies, July 1982, 49, 335–54. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 270 METHODOLOGY The Empirical Foundations of Calibration Hansen, Lars Peter, “Large Sample Properties of Generalized Method of Moments Estimators,” Econometrica, July 1982, 5 0, 1029–54. Hansen, Lars Peter, and Thomas J.Sargent, “Seasonality and Approximation Errors in Rational Expectations Models,” Journal of Econometrics, January/February 1993, 55, 21–55. Hansen, Lars Peter, and K.J.Singleton, “Stochastic Consumption, Risk Aversion, and the Temporal Behavior of Asset Returns,” Journal of Political Economy, 1983, 91:2, 249–65. Hansen, Lars Peter, Thomas J.Sargent, and T.D.Tallarini, “Pessimism and Risk in General Equilibrium,” manuscript, 1995. Harrison, J.Michael, and David M.Kreps, “Martingales and Arbitrage in Multiperiod Securities Markets,” Journal of Economic Theory, June 1979, 20, 381–408. Heaton, John, and D.Lucas, “Evaluating the Effects of Incomplete Markets on Risk Sharing and Asset Pricing,” Journal of Political Economy, forthcoming, June 1996. Heckman, James, and Richard Robb, Jr., “Alternative Methods for Evaluating the Impact of Interventions.” In Heckman, J., and Burton Singer, eds., Longitudinal Analysis of Labor Market Data. Cambridge, U.K.: Cambridge University Press, 1985, pp. 156–245. Heckman, James J., and Guilherme Sedlacek, “Heterogeneity, Aggregation, and Market Wage Functions: An Empirical Model of Self-Selection in the Labor Market,” Journal of Political Economy, December 1985, 93, 1077–125. Houthakker, Hendrik, “The Pareto Distribution and the Cobb-Douglas Production Function,” Review of Economic Studies, 1956, 23:1 27–31. G mrohorg u lu, Ayse, “The Welfare Costs of Inflation Under Imperfect Insurance,” Journal of Economic Dynamics and Control, January 1992, 16, 79–91. Klein, L.R., A Textbook of Econometrics. Evanston, Ill.: Row, Peterson and Company, 1953. Klein, L.A., and A.S.Goldberger, An Econometric Model of the United States, 1929– 52. Amsterdam: North-Holland, 1955. Kocherlakota, N., “The Equity Premium: It’s Still a Puzzle,” Journal of Economic Literature, forthcoming 1996. Kydland, Finn E., and Edward C.Prescott, “Time to Build and Aggregate 103 Fluctuations,” Econometrica, 1982, 5 0:6, 1345–70. Learner, E.E., Specification Searches: Ad Hoc Inference with Nonexperimental Data. New York: John Wiley & Sons, 1978. LeCam, L., and G.Yang, Asymptotics in Statistics: Some Basic Concepts. Berlin: SpringerVerlag, 1990. Liu, T.C., “Underidentification, Structural Estimation, and Forecasting,” Econometrica, October 1960, 28, 855–65. Lucas, Robert E., Jr., “Methods and Problems in Business Cycle Theory,” Journal of Money, Credit and Banking, November 1980, 12, 696–715; reprinted in Lucas, R.E., ed., Studies in Business Cycle Theory. Cambridge, Mass.: Massachusetts Institute of Technology Press, 1981, pp. 271–96. Lucas, Robert E., Jr., and Edward C.Prescott, “Investment Under Uncertainty,” Econometrica, 1971, 39:5, 659–81. Lucas, Robert E., Jr., and Leonard A.Rapping, “Real Wages, Employment, and Inflation,” Journal of Political Economy, September/October 1969, 77, 721–54. Lucas, Robert E., Jr., and Thomas J.Sargent, “After Keynesian Macroeconomics,” Quarterly Review, Federal Reserve Bank of Minneapolis, Spring 1979, 3, 1–16. MaCurdy, Thomas, “Econometric Model of the Labor Supply in a Life Setting,” Ph.D. dissertation, University of Chicago, 1978. MaCurdy, Thomas, “An Empirical Model of Labor Supply in a Life Cycle Setting,” Journal of Political Economy , 1981, 89:6, 1059–85. Mantel, Rolf R., “On the Characterization of Aggregate Excess Demand,” Journal of Economic Theory, March 1974, 7, 348–55. Mehra, Rajnish, and Edward C.Prescott, “The Equity Premium: A Puzzle,” Journal of Monetary Economics, March 1985, 15, 145–61. Morgan, M.S., The History of Econometrics Ideas. Cambridge: Cambridge University Press, 1990. Nerlove, M., Estimation and Identification of Cobb-Douglas Production Functions. Chicago: Rand McNally, 1965. Oreskes, Naomi, Kristen ShraderFrechette, and Kenneth Belitz, “Verification, Validation, and Confirmation of Numerical Models in the Earth Sciences,” Science, February 4, 1994, 263, 641–46. Peitgen, H.-O., and P.Richter, The Beauty of Fractals. Berlin: Springer-Verlag, 1986. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors EMPIRICAL FOUNDATIONS OF CALIBRATION 271 104 Journal of Economic Perspectives Prescott, Edward, “Theory Ahead of Business-Cycle Measurement,” Carnegie-Rochester Conference on Public Policy, 1986a, 25, 11– 44; reprinted in Quarterly Review, Federal Reserve Bank of Minneapolis, Fall 1986, 10, 9–22. Prescott, Edward, “Response to a Skeptic,” Quarterly Review, Federal Reserve Bank of Minneapolis, Fall 1986b, 10, 28–33. Ríos-Rull, José-Víctor, “Models with Heterogeneous Agents.” In Cooley, T.F., ed., Frontiers of Business Cycle Research. Princeton, N.J.: Princeton University Press, 1995, pp. 98–125. Rubinstein, Mark, “An Aggregation Theorem for Securities Markets,” Journal of Financial Economics, September 1974, 1, 225– 44. Sargent, Thomas J., “Two Models of Measurements, and the Investment Accelerator,” Journal of Political Economics, 1989, 97:2, 251–87. Sargent, Thomas J., and Christopher A.Sims, “Business Cycle Modeling Without Pretending to Have Too Much A Priori Economic Theory.” In Sims, Christopher, ed., New Methods in Business Cycle Research. Minneapolis: Federal Reserve Bank, 1977, pp. 45– 110. Scarf, Herbert, and T.Hansen, The Computation of Economic Equilibria. New Haven: Yale University Press, 1973. Shoven, John B., and John Whalley, “A General Equilibrium Calculation of the Effects of Differential Taxation of Income from Capital in the U.S.,” Journal of Public Economics, November 1972, 1, 281–322. Shoven, John B., and John Whalley, Applying General Equilibrium. New York: Cambridge University Press, 1992. Simon, Julian L., and Dennis J.Aigner, “Cross Sectional Budget Studies, Aggregate Time Series Studies and the Permanent In- come Hypothesis,” American Economic Review, June 1970, 60:2, 526–41. Sims, Christopher A., “Macroeconomics and Reality,” Econometrica, 1980, 48:1, 1–48. Sims, Christopher A., “Rational Expectations Modeling with Seasonally Adjusted Data,” Journal of Econometrics, January/February 1993, 55, 9–19. Smith, J., “A Comparison of the Earnings Patterns of Two Samples of JTPA Eligibles,” unpublished paper, University of Western Ontario, London, Canada, August 1995. Sonnenschein, H., “Do Walres Identity and Continuity Characterize the Class of Community Excess Demand Functions?,” Journal of Economic Theory, August 1973, 6, 345–54. Stoker, T.M., “Empirical Approaches to the Problem of Aggregation over Individuals,” Journal of Economic Literature, December 1993, 31:4, 1827–74. Taylor, John B., Macroeconomic Policy in a World Economy: From Econometric Design to Practical Operation. New York: W.W.Norton and Company, 1993. Tinbergen, J., Statistical Testing of Business Cycle Theories. Vol. 2, Business Cycles in the USA, 1919–1932, Geneva: League of Nations, 1939. Tobin, James, “A Statistical Demand Function for Food in the USA,” Journal of the Royal Statistical Society, 1950, 113, Series A, Part II, 113–41. Trostel, P.A., “The Effect of Taxation on Human Capital,” Econometrica, April 1993, 101, 327–50. Watson, M.W., “Measures of Fit for Calibrated Models,” Journal of Political Economy, 1993, 101:6, 1011–41. Wilson, Robert B., “The Theory of Syndicates,” Econometrica, January 1968, 36, 119–32. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 272 CHAPTER 15 Oxford Economic Papers 47 (1995), 24–44 FACTS AND ARTIFACTS: CALIBRATION AND THE EMPIRICAL ASSESSMENT OF REAL-BUSINESS-CYCLE MODELS By KEVIN D.HOOVER Department of Economics, University of California, Davis, California 95616–8578, USA 1. Whither quantitative macroeconomics? THE RELATIONSHIP between theory and data has been, from the beginning, a central concern of the new-classical macroeconomics. This much is evident in the title of Robert E.Lucas’s and Thomas J.Sargent’s landmark edited volume, Rational Expectations and Econometric Practice (1981). With the advent of real-business-cycle models, many new classical economists have turned to calibration methods. The new classical macroeconomics is now divided between calibrators and estimators. But the debate is not a parochial one, raising, as it does, issues about the relationships of models to reality and the nature of econometrics that should be important to every school of macroeconomic thought, indeed to all applied economics. The stake in this debate is the future direction of quantitative macroeconomics. It is, therefore, critical to understand the root issues. Lucas begins the second chapter of his Models of Business Cycles with the remark: Discussions of economic policy, if they are to be productive in any practical sense, necessarily involve quantitative assessments of the way proposed policies are likely to affect resource allocation and individual welfare. (Lucas 1987, p. 6; emphasis added) This might appear to be a clarion call for econometric estimation. But appearances are deceiving. After mentioning Sumru Altug’s (1989) estimation and rejection of the validity of a variant of Finn E.Kydland and Edward C.Prescott’s (1982) real-businesscycle model (a model which takes up a large portion of his book), Lucas writes: …the interesting question is surely not whether [the real-business-cycle model] can be accepted as ‘true’ when nested within some broader class of models. Of course the model is not ‘true’: this much is evident from the axioms on which it is constructed. We know from the onset in an enterprise like this (I would say, in any effort in positive economics) that what will emerge–at best–is a workable approximation that is useful in answering a limited set of questions. (Lucas 1987, p. 45) Lucas abandons not only truth but also the hitherto accepted standards of empirical economics. Models that clearly do not fit the data, he argues, may nonetheless be calibrated to provide useful quantitative guides to policy. Calibration techniques are commonly applied to so-called ‘computable general-equilibrium’ models. They were imported into macroeconomics as a © Oxford University Press 1995 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors FACTS AND ARTIFACTS K.D.HOOVER 273 25 means of quantifying real-business-cycle models, but now have a wide range of applications. Some issues raised by calibration are common to all computable general-equilibrium models; the concern of this paper, however, is with realbusiness-cycle models and related macroeconomic applications; and, as will appear presently, these raise special issues. A model is calibrated when its parameters are quantified from casual empiricism or unrelated econometric studies or are chosen to guarantee that the model precisely mimics some particular feature of the historical data. For example, in Kydland and Prescott (1982), the coefficient of relative risk aversion is justified on the basis of microeconomic studies, while the free parameters of the model are set to force the model to match the variance of GNP without any attempt to find the value of empirical analogues to them. Allan W.Gregory and Gregor W.Smith (1991, p. 3) conclude that calibration ‘…is beginning to predominate in the quantitative application of macroeconomic models’. While indicative of the importance of the calibration methodology, Gregory and Smith’s conclusion is too strong. Aside from the new classical school, few macroeconomists are staunch advocates of calibration. Within the new classical school, opinion remains divided. Even with reference to real-business-cycle models, some practitioners have insisted that calibration is at best a first step, which must be followed ‘…by setting down a metric (e.g. one induced by a likelihood function) and estimating parameters by finding values that make the metric attain a minimum’ (Gary Hansen and Sargent 1988, p. 293).1 Sargent advocates estimation or what Kydland and Prescott (1991) call the ‘system-of-equations approach’. Estimation has been the standard approach in macroeconometrics for over 40 years. Sargent and like-minded new classical economists modify the standard approach only in their insistence that the restrictions implied by dynamic-optimization models be integrated into the estimations. The standard of empirical assessment is the usual one: how well does the model fit the data statistically? Lucas and Kydland and Prescott reject statistical goodness of fit as a relevant standard of assessment. The issue at hand might then be summarized: who is right–Lucas and Kydland and Prescott, or Sargent? The answer to this question is not transparent. Estimation is the status quo. And, although enthusiastic advocates of calibration already announce its triumph, its methodological foundations remain largely unarticulated. An uncharitable interpretation of the calibration methodology might be that the advocates of real-business-cycle models are so enamored of their creations that they would prefer to abandon commonly accepted, neutral standards of empirical evaluation (i.e. econometric hypothesis testing) to preserve their 1 Despite the joint authorship of the last quotation, I regard Sargent and not Hansen as the preeminent proponent of the necessity of estimation, because I recall him forcefully insisting on it in his role as discussant of a paper by Thomas F.Cooley and Hansen (1989) at the Federal Reserve Bank of San Francisco’s Fall Acacemic Conference; see also Manuelli and Sargent (1988, pp. 531–4). © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 274 METHODOLOGY 26 FACTS AND ARTIFACTS models. This would be an ad hoc defensive move typical of a degenerating research program. This interpretation is not only uncharitable, it is wrong. Presently, we shall see that Herbert Simon’s (1969) Sciences of the Artifical provides the materials from which to construct a methodological foundation for calibration, and that calibration is compatible with a well-established approach to econometrics that is nonetheless very different from the Cowles Commission emphasis on the estimation of systems of equations. Before addressing these issues, however, it will be useful to describe the calibration methodology and its place in the history and practice of econometrics in more detail. 2. The calibration methodology 2.1. The paradigm case Kydland and Prescott (1982) is the paradigm new-classical equilibrium, realbusiness-cycle model. It is neoclassical optimal-growth model with stochastic shocks to technology which cause the equilibrium growth path to fluctuate about its steady state.2 Concrete functional forms are chosen to capture some general features of business cycles. Production is governed by a constant-elasticity-ofsubstitution production function in which inventories, fixed capital, and labor combine to generate a single homogeneous output that may either be consumed or reinvested. Fixed capital requires a finite time to be built before it becomes a useful input. The constant-relative-risk-aversion utility function is rigged to possess a high degree of intertemporal substitutability of leisure. Shocks to technology are serially correlated. Together the structure of the serial correlation of the technology shocks and the degree of intertemporal substitution in consumption and leisure choices govern the manner in which shocks are propagagated through time and the speed of convergence back towards the steady state. Once the model is specified, the next step is to parameterize its concrete functional forms. Most of the parameters of the model are chosen from values culled from other applied econometric literatures or from general facts about national-income accounting. For example, Thomas Mayer (1960) estimated the average time to construct complete facilities to be 21 months; Robert E.Hall (1977) estimated the average time from start of projects to beginning of production to be two years. Citing these papers, but noting that consumer durable goods take considerably less time to produce, Kydland and Prescott (1982, p. 1361) assume that the paramters governing capital formation are set to imply steady construction over four quarters.3 The values for depreciation rates and the capital/inventory ratio are set to rough averages from the national2 For general descriptions of the details and varieties of real-business-cycle models see Lucas (1987), Kevin D.Hoover (1988, ch. 3), and Bennett T.McCallum (1989). Steven M.Sheffrin (1989, ch. 3, especially pp. 80, 81), gives a step-by-step recipe for constructing a prototypical real-business-cycle model. 3 Mayer’s estimates were for complete projects only, so that new equipment installed in old plants, which must have a much shorter time-to-build than 21 months, was not counted. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors FACTS AND ARTIFACTS K.D.HOOVER 275 27 income accounts. Ready estimates from similar sources were not available for the remaining six paramters of the model, which include parameters governing intertemporal substitution of leisure and the shocks to technology. These were chosen by searching over possible parameter values for a combination that best reproduced certain key variances and covariances of the data. In particular, the technology shock variance was chosen in order to exactly match the variance of output in the postwar US economy. To test the model’s performance, Kydland and Prescott generate a large number of realizations of the technology shocks for 118 periods corresponding to their postwar data. They then compute the variances and covariances implied by the model for a number of important variables: output, consumption, investment, inventories, the capital stock, hours worked, productivity, and the real rate of interest.4 These are then compared with the corresponding variances and covariances of the actual US data.5 Kydland and Prescott offer no formal measure of the success of their model. They do note that hours are insufficiently variable with respect to the variability of productivity to correspond accurately to the data, but otherwise they are pleased with the model’s ability to mimic the second moments of the data. Real-business-cycle models, treated in the manner of Kydland and Prescott, are a species of the genus computable (or applied) general-equilibrium models. The accepted standards for implementing computable generalequilibrium models, as codified, for example, in Ahsan Mansur and John Whalley (1984), do not appear to have been adopted in the real-businesscycle literature. For example, while some practitioners of computable generalequilibrium models engage in extensive searches of the literature in order to get some measure of the central tendency of assumed elasticities, Kydland and Prescott’s (1982) choice of parameterization appears almost casual. Similarly, 4 In fact, it is not clear in Kydland and Prescott (1982) that these are calculated from the cross-section of a set of realizations or from a single time-series realization. In a subsequent paper that extends their results, Kydland and Prescott (1988, p. 353) are quite precise about using a cross-section of many realizations. Because they are interested only in the moments of the variables and not in particular time-series, Kydland and Prescott initialize variables to their steady-state values or, equivalently in the context of detrended data, to zero. In order to generate a time path that can be compared to the history of a particular series, it is necessary, as in Hansen and Prescott (1993), to initialize at some actual historical benchmark. 5 Despite my referring to Kydland and Prescott’s model as a growth model, the model for which they calculate the variances and covariances does not possess an exogenous source of trend growth. Thus, to make comparisons, Kydland and Prescott (1982, p. 1362) detrend the actual data using the Hodrick-Prescott filter. The particular choice of filter is not defended in any detail. Prescott (1983, p. 6) simply asserts that it produces ‘about the right degree of smoothness, when fit to the logarithm of the real GNP series’ without any indication by what standard rightness is to be judged. Kydland and Prescott (1990, p. 9) claim that it generates a trend close to the trend that students of business cycles would draw by hand through a plot of actual GNP. Although the Hodrick-Prescott filter is almost universally adopted in comparing real-business-cycle models to actual data, Fabio Canova (1991b) shows that the use of Hodrick-Prescott filters with different choices for the values of a key parameter or of several entirely different alternative filters radically alters the cyclical characteristics of economic data (also see Timothy Cogley and James Nason, 1993). © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 276 28 METHODOLOGY FACTS AND ARTIFACTS although Kydland and Prescott report some checks on robustness, these appear to be perfunctory.6 In the context of computable general-equilibrium models, calibration is preferred in those cases in which, because of extensive disaggregation, the number of parameters is too large relative to the available data set to permit econometric estimation.7 Since typical real-business-cycle models are one-good, one-agent models, there is no difficulty in estimating them using standard methods such as maximum likelihood or generalized method of moments. Indeed, since the practitioners are often interested principally in matching selected second moments, method-of-moments estimators can concentrate on the moments of interest to the exclusion of others (see Watson 1993, p. 1037). As noted earlier, Altug (1989) estimates and rejects a close relative of Kydland and Prescott’s model using maximum likelihood. The central problem of this paper can be restated: is there a case for ignoring Altug’s rejection of the Kydland and Prescott model? The case must be something other than the standard one of too many parameters found in the literature on computable general-equilibrium models. 2.2. Calibration as estimation Various authors have attempted to tame calibration and return it to the traditional econometric fold. Manuelli and Sargent (1988), Gregory and Smith (1990a), Canova (1991a), and Bansal et al. (1991) interpret calibration as a form of ‘estimation by simulation’. In such a procedure, parameters are chosen, and the relevant features of the simulated output of the calibrated model are compared to the analogous features of the actual data. Such a procedure differs from standard estimation methods principally in that it allows the investigator to expand or restrict the range of features considered to be relevant. Lucas’s argument, however, is that any form of estimation is irrelevant. In their early writings, Kydland and Prescott were not in fact as explicit as Lucas about the irrelevance of estimation. They merely argued that it would be premature to apply techniques to their model, such as those developed by Lars Peter Hansen and Sargent (1980), to account for the systemic effects of rational expectations (Kydland and Prescott, 1982, p. 1369). Prescott (1983, pp. 8–11) was more pointed: real-business-cycle models are tightly parameterized. They 6 Canova (1991a) suggests a formal methodology and provides an example in which sensitivity analysis is conducted with respect to distributions for parameter values constructed from the different values reported in unrelated studies or from a priori information on the practically or theoretically admissible range of parameter values. 7 Lawrence J.Lau (1984) notices that any model that can be calibrated can also be estimated. He uses ‘calibration’, however, in a narrow sense. A model is calibrated when its parameters are chosen to reproduce the data of a benchmark period. Thus, parameterization on the basis of unrelated econometric studies does not count as calibration for him. Lau’s usage is diametrically opposed to that of Gregory and Smith (1991) for whom calibration is only the assignment of parameter values from unrelated sources. We use ‘calibration’ in both Lau’s and Gregory and Smith’s senses. Lau and, similarly, James MacKinnon (1984) make strong pleas for estimation instead of, or in addition to, calibration, and for subjecting computable general-equilibrium models to statistical specification tests. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors FACTS AND ARTIFACTS 277 K.D.HOOVER 29 will almost inevitably be rejected against a weakly restricted alternative hypothesis, but such alternative hypotheses arise from the introduction of arbitrary stochastic processes and, so, are not suitable benchmarks for economic inference. ‘A model may mimic the cycle well but not perfectly’ (Prescott 1983, p. 10). Similarly, Kydland and Prescott (1991, p. 174) write: Unlike the system-of-equations approach, the model economy which better fits the data is not [necessarily?] the one used. Rather, currently established theory dictates which one is used. The dominance of theory in the choice of models lies at the heart of the difference between estimators and calibrators. To throw the difference into high relief, one can think of estimators pursuing a competitive strategy and calibrators pursuing an adaptive strategy. Under the competitive strategy, theory proposes, estimation and testing disposes. In fine, alternative theories compete with one another for the support of the data. The adaptive strategy begins with an unrealistic model, in the sense of one that is an idealized and simplified product of the core theory. It sees how much mileage it can get out of that model. Only then does it add any complicating and more realistic feature. Unlike the competitive strategy, the aim is never to test and possibly reject the core theory, but to construct models that reproduce the economy more and more closely within the strict limits of the basic theory. The distinction between the competitive and adaptive strategies is sharply drawn and somewhat stylized, but focuses nonetheless on a genuine difference. On the one hand, the competitive strategy is the received view of econometricians, taught in an idealized form in most econometric textbooks, even if more honored in the breach than the observance by applied economists. The competitive strategy is explicit in Gregory and Smith’s (1990b) ‘Calibration as Testing’. Even if in practice no theory is ever decisively rejected through a test based on an econometic estimation, the theory is nonetheless regarded as at risk and contingent—even at its core. On the other hand, the real-business-cycle modeller typically does not regard the core theory at risk in principle. Like the estimators, the calibrators wish to have a close fit between their quantified models and the actual data—at least in selected dimensions. But the failure to obtain a close fit (statistical rejection) does not provide grounds for rejecting the fundamental underlying theory. Adaptation in the face of recalcitrant data is adaptation of peripheral assumptions, not of the core. Thus, the inability of Kydland and Prescott’s (1982) original real-business-cycle model to match the data prompted more complicated versions of essentially the same model that included, for example, heterogeneous labor (Kydland 1984), a banking sector (Robert G.King and Charles I.Plosser 1984), indivisible labor (Gary Hansen 1985), separate scales for straight-time and overtime work (Gary Hansen and Sargent 1988), and variable capital intensity (Kydland and Prescott 1988). One consequence of these strategies is that esimators possess a common ground, the performance of each theoretically-based specification against actual © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 278 METHODOLOGY 30 FACTS AND ARTIFACTS data, on which to judge the performance of competing models. For the calibrators, however, data help discriminate only b etween different adaptations of the common core. The core theory itself is not questioned, so that, unintentially perhaps, the core theory becomes effectively a Lakatosian hardcore (Lakatos 1970, 1978; Blaug 1992, ch. 2). Calibration does not provide a method that could in principle decide between fundamentally different business-cycle models (e.g. real-business-cycle models or Keynesian business-cycle models) on the basis of empirical evidence derived from the calibration exercise itself.8 Critics of real-business-cycle models who attempt such comparisons fall back either on attacking the discriminating power of calibration methods (e.g. Hartley et al. 1993) or on adaptations of standard econometric techniques (e.g. Canova et al. 1993). Kydland and Prescott are explicit in rejecting these applications of estimation techniques as missing the point of the calibration method. The aim of this paper is partly to explicate and appraise their view.9 2.3. The mantel of Frisch Calibrators radically reject the system-of-equations approach. But Kydland and Prescott, at least, do not reject econometrics. Rather, they argue that econometrics is not coextensive with estimation; calibration is econometrics. Kydland and Prescott (1991, pp. 161, 162) point out that for Ragnar Frisch, Irving Fisher, and Joseph Shumpeter, the founders of the Econometric Society, ‘econometrics’ was the unification of statistics, economic theory, and mathematics. Unacknowledged by Kydland and Prescott, Mary Morgan’s (1990) brilliant history of econometrics supports and elaborates their point. According to Morgan, even before the term ‘econometrics’ had wide currency, the econometric ideal had been to weld mathematical, deductive economics to statistical, empirical economics to provide a substitute for the experimental methods of the natural sciences appropriate to the study of society. This ideal collapsed with the rise of the system-of-equations approach in the wake of the Cowles Commission. Kydland and Prescott point to Frisch’s (1933) article, ‘Propagation Problems and Impulse Response Problems in Dynamic Economics’ as a precursor to both their own real-business-cycle model and to calibration methods. Frisch argues that quantitative analysis requires complete models: i.e. general-equilibrium models in a broad sense. He considers a sequence of models, starting with a very simple one, and then adding complications. He models the time-to-build feature of capital formation. He distinguishes between the impulses that start business cycles and the dynamic mechansisms that amplify and propagate them. He quantifies his models using calibration techniques. And, precisely like Kydland and Prescott 8 This is not to say that there could not be some other basis for some decision. Hoover 1994a (as well as work in progress) outlines a possible method of using econometric techniques in a way that respects the idealized nature of the core models without giving up the possibility of empirical discrimination. 9 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors FACTS AND ARTIFACTS 279 K.D.HOOVER 31 (1982), Frisch marvels at how well a very simple model can capture the features of actual data. Although Kydland and Prescott are correct to see the affinities between Frisch’s work and their own, they ignore the very real differences between Frisch and themselves. Frisch’s approach is wholly macroeconomic. Frisch writes: In order to attack these problems on a macro-dynamic basis so as to explain the movement of the system taken in its entirety, we must deliberately disregard a considerable amount of the details of the picture. We may perhaps start by throwing all kinds of production into one variable, all consumption into another, and so on, imagining that the notions ‘production’, ‘consumption’, and on, can be measured by some sort of total indices. (1933, p. 173) While his flight to aggregates parallels the practice of the new classical realbusiness-cycle model, Frisch does not suggest that this is a way station on the road to microfoundations. His article does not hint at the desirability of microfoundations, even of the pseudo-microfoundations of the representativeagent model: there is not an optimization problem to be found. Frisch appears to use calibration mainly for purposes of illustration, and not to advocate it as a preferred technique. He writes: At present I am guessing very roughly at these parameters, but I believe that it will be possible by appropriate statistical methods to obtain more exact information about them. I think, indeed, that the statistical determination of such structural parameters will be one of the main objectives of the economic cycle analysis of the future. (1933, p. 185) Precisely which statistical methods are appropriate appears to be an open question.10 More generally, although Frisch stresses the importance of theory, there is no hint that his interpretation is limited to ‘maximizing behavior subject to constraints’ (Kydland and Prescott 1991, p. 164). Frisch does not define ‘theory’ in ‘Propagation Problems…’, but the examples he produces of theories are not of an obviously different character from the structures employed by Jan Tinbergen, Lawrence Klein, James Dusenberry, and the other ‘Keynesian’ macromodelers who are the special bugbears of the advocates of new-classical, real-business-cycle models. Schumpeter (co-founder with Frisch of the Econometric Society) provides typically prolix discussions of the meaning of ‘economic theory’ in his magisteral History of Economic Analysis (1954). For Schumpeter (1954, pp. 14, 15), theories 10 Frisch’s own shifting views illustrate how open a question this was for him. By 1936, he had backtracked on the desirability of estimating calibrated models. In 1938, he argued that structural estimation was impossible because of pervasive multicollinearity. In its place he proposed estimating unrestricted reduced forms (see Morgan, 1990, p. 97; also see Aldrich 1989, section 2). In this he comes much closer to Christopher Sims (1980) program of vector autoregressions without ‘incredible’ identifying restrictions. I am grateful to an anonymous referee for reminding me of this point. Hoover (1992) identifies Sims’s program as one of three responses to the Lucas (1976) critique of policy invariance. Calibration and the systems-of-equations approach each possess a analogous response. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 280 32 METHODOLOGY FACTS AND ARTIFACTS are, on the one hand, ‘synonomous with Explanatory Hypotheses’, and on the other hand, ‘the sum total of the gadgets’, such as ‘“marginal rate of substitution”, “marginal productivity”, “multiplier”, “accelerator”’, including ‘stragegically useful assumptions’, ‘by which results may be extracted from the hypothesis’. Schumpeter concludes: ‘In Mrs. Robinson’s unsurpassingly felicitous phrase, economic theory is a box of tools’. Later Schumpeter defends the theoretical credentials of Wesley C.Mitchell, the subject of Tjalling Koopman’s (1947) famous attack on ‘Measurement without Theory’: …in intention as well as in fact, he was laying the foundations for a ‘theory’, a business cycle theory as well as a general theory of the economic process, but for a different one. (1954, p. 1166) Kydland and Prescott (1991, p. 164) argue that the system-of-equations approach flourished in the 1950s only because economists lacked the tools to construct stochastic computable general-equilibrium models. They proclaim the death of the system-of-equations approach: The principal reason for the abandonment of the system-of-equations approach, however, was the advances in neoclassical theory that permitted the application of the paradigm in dynamic stochastic settings. Once the neoclassical tools needed for modeling business cycle fluctuations existed, their application to this problem and their ultimate domination over any other method was inevitable. (1991, p. 167) This is an excessively triumphalist and whiggish history of the development of econometric thought. First, the work of Frisch and others in the 1930s provides no support for Kydland and Prescott’s narrowing of the meaning of ‘theory’ to support such tendentious statements as: ‘To summarize the Frisch view, then, econometrics is quantitative neoclassical theory with a basis in facts’ (Kydland and Prescott 1991, p. 162; emphasis added). (A model is ‘neoclassical’ for Kydland and Prescott (1991, p. 164) when it is constructed from ‘…agents maximizing subject to constraints and market clearing’.) Second, the declaration of the death of the system-of-equations approach is premature and greatly exaggerated. Allegiance to the system-of-equations approach motivates the many efforts to interpret calibration as a form of estimation. Third, the calibration methodology is not logically connected to Kydland and Prescott’s preferred theoretical framework. The example of Frisch shows that calibration can be applied to models that are not stochastic dynamic optimal-growth models. The example of Lars Peter Hansen and Sargent (1980) shows that, even those who prefer such models, can use them as the source of identification for systems of equations— refining rather than supplanting the traditional econometrics of estimation. 3. The quantification of theory Although Kydland and Prescott overstate the degree to which Frisch and the econometrics of the 1930s foreshadowed their work, they are nonetheless correct to note many affinities. But such affinities, even if they were more complete than they turn out to be, do not amount to an argument favoring calibration © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors FACTS AND ARTIFACTS 281 K.D.HOOVER 33 over estimation. At most, they are an illicit appeal to authority. To date, no compelling defence of the calibration methodology has been offered. An interpretation of the point of calibration and an assessment of its merits can be constructed, however, from hints provided in Lucas’s methodological writings of the 1970s and early 1980s. 3.1. Models ‘Model’ is a ubiquitous term in economics, and a term with a variety of meanings. One commonly speaks of an econometric model. Here one means the concrete specification of functional forms for estimation. I call these observational models. The second main class of models are evaluative or interpretive models. An obvious subclass of interpretive/evaluative models are toy models. A toy model exists merely to illustrate or to check the coherence of principles or their interaction. An example of a toy model is the overlappinggenerations model with money in its simplest incarnations. No one would think of drawing quantitative conclusions about the working of the economy from it. Instead one wants to show that models constructed on its principles reproduce certain known qualitative features of the economy and suggest other qualitative features that may not have been known or sufficiently appreciated (cf. Diamond 1984, p. 47), Were one so rash as to estimate such a model, it would surely be rejected, but that would be no reason to abandon it as a testbed for general principles. Is there another subclass of interpretive/evaluative models, one that involves quantification? Lucas seems to think so: One of the functions of theoretical economics is to provide fully articulated, artificial economic systems that can serve as laboratories in which policies that would be prohibitively expensive to experiment within actual economies can be tested out at much lower cost. (Lucas 1980, p. 271) Let us call such models benchmark models. Benchmark models must be abstract enough and precise enough to permit incontrovertible answers to the questions put to them. Therefore, …insistence on the ‘realism’ of an economic model subverts its potential usefulness in thinking about reality. Any model that is well enough articulated to give clear answers to the questions we put to it will necessarily be artificial, abstract, patently unreal. (Lucas 1980, p. 271) On the other hand, only models that mimic reality in important respects will be useful in analyzing actual economies. The more dimensions in which the model mimics the answers actual economies give to simple questions, the more we trust its answers to harder questions. This is the sense in which more ‘realism’ in a model is clearly preferred to less. (Lucas 1980, p. 272) Later in the same essay, Lucas emphasizes the quantitative nature of such model building: © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 282 34 METHODOLOGY FACTS AND ARTIFACTS Our task…is to write a FORTRAN program that will accept specific economic policy rules as ‘input’ and will generate as ‘output’ statistics describing the operating characteristics of time series we care about, which are predicted to result from these policies, (p. 288) For Lucas, Kydland and Prescott’s model is precisely such a program.11 One might interpret Lucas’s remarks as making a superficial contribution to the debate over Milton Friedman’s ‘Methodology of Positive Economics’ (1953): must the assumptions on which a theory is constructed be true or realistic or is it enough that the theory predicts ‘as if’ they were true? But this would be a mistake. Lucas is making a point about the architecture of models and not about the foundations of secure prediction. Lucas refers to a model as fully ‘realistic’ when it fully accounts for all the factors that determine the variables of interest. Lucas makes two points. Uncontroversially, he argues that toy models convey deep understanding of economic principles. More interestingly, he argues that benchmark models have an advantage over estimation. This is controversial because estimators believe that fully articulated specifications are required for accurate quantification. This is expressed in their concern for specification error, omitted variable bias, and so forth. Their view is widely shared. The point is not that estimated models are necessarily more realistic in Lucas’s sense than calibrated models, nor that estimation is the only or even the most reliable way to quantify a model or its components.12 Rather it is that any method of quantification that does not aspire to full articulation is likely to mislead. Lucas denies this, and the interesting issues are how to appraise his position, and, if his position is sustainable, how to appraise quantified benchmark models themselves. To make this clear, consider Lucas’s (1987, pp. 20–31) cost-benefit analysis of the policies to raise GNP growth and to damp the business cycle. Lucas’s model considers a single representative consumer with a constant-relative-risk-aversion utility function facing an exogenous consumption stream. The model is calibrated by picking reasonable values for the mean and variance of consumption, the subjective rate of discount, and the constant coefficient of relative risk aversion. Lucas then calculates how much extra consumption consumers would require to compensate them in terms of utility for a cut in the growth of consumption and how much consumption they would be willing to give up to secure smoother consumption streams. Although the answers that Lucas seeks are quantitative, the model is not used to make predictions that might be subjected to statistical tests. Indeed, it is a striking illustration of why calibration should not be interpreted as estimation by simulation. Lucas’s model is used to set upper bounds to the benefits that might conceivably be gained in the real world. Its parameters must reflect some truth about the world if it is to 11 Kydland and Prescott do not say, however, whether it is actually written in FORTRAN. For example, to clarify a point raised by an anonymous referee, if the central bank had direct knowledge of it money supply function, that would be better than estimating it. 12 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors FACTS AND ARTIFACTS 283 K.D.HOOVER 35 be useful, but they could not be easily directly estimated. In that sense, the model is unrealistic.13 3.2. Artifacts In a footnote, Lucas (1980, p. 272, fn. 1) cites Simon’s Sciences of the Artificial (1969) as an ‘immediate ancestor’ of his ‘condensed’ account. To uncover a more fully articulated argument for Lucas’s approach to modelling, it is worth following up the reference. For Simon, human artifacts, among which he must count economic models, can be thought of as a meeting point—an ‘interface’…—between an ‘inner’ environment, the substance and organization of the artifact itself, and an ‘outer’ environment, the surroundings in which it operates. (Simon 1969, pp. 6, 7) An artifact is useful, it achieves its goals, if its inner environment is appropriate to its outer environment. The distinction between the outer and inner environments is important because there is some degree of independence between them. Clocks tell time for the outer environment. Although they may indicate the time in precisely the same way, say with identical hands on identical faces, the mechanisms of different clocks, their inner environments, may be constructed very differently. For determining when to leave to catch a plane, such differences are irrelevant. Equally, the inner environments may be isolated from all but a few key features of the outer environment. Only light entering through the lens for the short time that its shutter is open impinges on the inner environment of the camera. The remaining light is screened out by the opaque body of the camera, which is an essential part of its design. Simon factors adaptive systems into goals, outer environments and inner environments. The relative independence of the outer and inner environments means that [w]e might hope to characterize the main properties of the system and its behavior without elaborating the detail of either the outer or the inner environments. We might look toward a science of the artificial that would depend on the relative simplicity of the interface as its primary source of abstraction and generality. (Simon 1969, p. 9) Simon’s views reinforce Lucas’s discussion of models. A model is useful only if it foregoes descriptive realism and selects limited features of reality to reproduce. The assumption upon which the model is based do not matter, so long as the model succeeds in reproducing the selected features. Friedman’s ‘as if’ methodology appears vindicated. 13 Of course, Lucas’s approach might be accepted in principle and still rejected in detail. For example, McCallum (1986, pp. 411, 412) objects to the characterization of consumption as fluctuating symmetrically about trend that is implicit in Lucas’s use of a mean/variance model. If the fluctuations of consumption are better described as varying degrees of shortfall relative to the trend of potential maximum consumption, then the benefits of consumption smoothing will be considerably higher than Lucas’s finds. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 284 36 METHODOLOGY FACTS AND ARTIFACTS But this is to move too fast. The inner environment is only relatively independent of the outer environment. Adaptation has its limits. In a benign environment we would learn from the motor only what it had been called upon to do; in a taxing environment we would learn something about its internal structure—specifically, about those aspects of the internal structure that were chiefly instrumental in limiting performance. (Simon 1969, p. 13)14 This is a more general statement of principles underlying Lucas’s (1976) critique of macroeconometric models. A benign outer environment for econometric models is one in which policy does not change. Changes of policy produce structural breaks in estimated equations: disintegration of the inner environment of the models. Economic models must be constructed like a ship’s chronometer, insulated from the outer environment so that ‘…it reacts to the pitching of the ship only in the negative sense of maintaining an invariant relation of the hands on its dial to real time, independently of the ship’s motions’ (Simon 1969, p. 9). Insulation in economic models is achieved by specifying functions whose parameters are invariant to policy. The independence of the inner and outer environments is not something which is true of arbitrary models; rather it must be built into models. While it may be enough in hostile environments for models to reproduce key features of the outer environment ‘as if’ reality was described by their inner environments, it is not enough if they can do this only in benign environments. Thus, for Lucas, the ‘as if’ methodology interpreted as an excuse for complacency with respect to modeling assumptions must be rejected. Simon’s notion of the artifact helps justify Lucas’s both rejecting realism in the sense of full articulation and at the same time, insisting that only through carefully constructing the model from invariants—tastes and technology, in Lucas’s usual phrase—can the model secure the benefits of a useful abstraction and generality. Recognizing that a model must be constructed from invariants does not itself tell us how to quantify it. The emphasis on a maintained theory or inner environment presents a generic risk for quantified idealized models (see Section 2.2 above). The risk is particularly severe for the calibration methodology with its adaptive strategy. Gregory and Smith (1991, p. 30) observe that ‘[s]etting parameter values (i.e. calibrating), simulating a model and comparing properties of simulations to those of data often suggests fruitful modifications of the model’. Generally, such modifications leave the essential core theory intact and attempt to better account for the divergences from the ideal, to better account for the fudge factors need to link the output of the model to the phenomenal laws. The risk, then, is that the core of the model becomes completely insulated from empirical confirmation or disconfirmation—even in the weakest senses of those terms. Kydland and 14 Haavelmo (1944, p. 28) makes a similar point in his well-known example of the failure of autonomy: the relationship between the speed of a car and the amount of throttle may be well-defined under uniform conditions, but would break down immediately the car was placed in a different setting. To understand how the car will perform on the track as well as on the road requires us to repair to the deeper principles of its operation. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors FACTS AND ARTIFACTS K.D.HOOVER 285 37 Prescott (1991, p. 171) explicitly deny that the confidence in the answers a model gives to policy questions can ‘…be resolved by computing some measure of how well the model economy mimics historical data’. Rather, ’[t]he degree of confidence in the answer depends on the confidence that is placed in the economic theory being used’. Kydland and Prescott do not explain what alternative sources there might be to justify our confidence in theory; the adaptive strategy of the calibration approach almost guarantees that empirical evidence will not be among those sources. 3.3. Quantification without history Calibrators of real-business-cycle models typically concentrate on matching selected second moments of variables rather than, say, matching the actual historical evolution of the modeled variables. Why? Lucas (1977, p. 218) observes that ‘business cycles are all alike’, not in exact detail but qualitatively. An informative test of a model’s ability to capture business-cycle behavior is not, therefore, its fit to some historical time series, which is but one of many possible realizations, but rather its ability to characterize the distribution from which that realization was drawn. Lucas (1977, pp. 219, 234) advocates the test of Irma Adelman and Frank L.Adelman (1959). The Adelmans asked the question, could one distinguish data generated by simulating a model (in their case, the Klein-Goldberger macroeconometric model) from actual data describing the economy, in the absence of knowledge of which was which? The Adelmans’ test compares the distribution of outcomes of the model to the actual economy. Once a close relation is established, to experiment with alternative policy rules is an easy next step. Even though government is not modelled in Kydland and Prescott’s initial models, policy analysis is their ultimate goal (Kydland and Prescott, 1982, p. 1369). Concentration on the second moments of variables can be seen as the practical implementation of the Adelmans’ standard: one eschews the particular realization in favor of a more general characterization of the distribution of possible outcomes.15 One reason, therefore, not to apply a neutral statistical test for the match between model and reality is that it is along only selected dimensions that one cares about the model’s performance at all. This is completely consistent with Simon’s account of artifacts. New classical economics has traditionally been skeptical about discretionary economic policies. New classical economists are, therefore, more concerned to evaluate the operating characteristics of policy rules. For this, the fit of the model to a particular historical realization is largely irrelevant, unless it assures it will also characterize the future distribution of 15 The Adelmans themselves examine the time-series properties of a single draw, rather than the characteristics of repeated draws. This probably reflects, in part, the computational expense of simulating a large macroeconometric model with the technology of 1959. It also reflects the application of Burns and Mitchell’s techniques for characterizing the repetitive features of business cycles through averaging over historical cycles all normalized to a notional cycle length. King and Plosser (1989) attempt to apply precisely Burns and Michell’s techniques to outcomes generated by a real-business-cycle model. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 286 38 METHODOLOGY FACTS AND ARTIFACTS outcomes. The implicit claim of most econometrics is that it does assure a good characterization. Probably most econometricians would reject calibration methods as coming nowhere close to providing such assurance. Substantial work remains to be done in establishing objective, comparative standards for judging competing models. 4. Aggregation and general equilibrium16 Whether calibrated or estimated, real-business-cycle models are idealizations along many dimensions. The most important dimension of idealization is the the models deal in aggregates while the economy is composed of individuals. After all, the distinction between microeconomics and macroeconomics is the distinction between the individual actor and the economy as a whole. All new classical economists believe that one understands macroeconomic behavior only as an outcome of individual rationality. Lucas (1987, p. 57) comes close to adopting the Verstehen approach of the Austrians.17 The difficulty with this approach is that there are millions of people in the economy and it is not practical—nor is it ever likely to become practical—to model the behavior of each of them.18 Universally, new classical economists adopt representativeagent models, in which one agent or a few types of agents, stand in for the behavior of all agents.19 The conditions under which a single agent’s behavior can accurately represent the behavior of an entire class are onerous and almost certainly never fulfilled in an actual economy. One interpretation of the use of calibration methods in macroeconomics is that the practitioners recognize that highly aggregated theoretical models must be descriptively false, so that estimates of them are bound to fit badly in comparison to atheoretical econometric models, which are able to exploit large numbers of free parameters. The theoretical models are nonetheless to be preferred because policy evaluation is possible only within their structure. In this, they are exactly like Lucas’s benchmark consumption model (see Section 3.1, above). Calibractors appeal to microeconomic estimates of key parameters because information about individual agents is lost in the aggregatation process. Estimators, in contrast, could argue that the idealized representative-agent 16 Aggregation and the problems it poses for macroeconomics are the subject of a voluminous literature. The present discussion is limited to a narrow set of issues most relevant to the question of idealization. 17 For a full discussion of the relationship between new classical and Austrian economics see Hoover (1988, ch. 10). 18 In Hoover (1984, pp. 64–6; and 1988, pp. 218–20), I refer to this as the ‘Cournot problem’ since it was first articulated by Augustin Cournot ([1838] 1927, p. 127). 19 Some economists reserve the term ‘representative-agent models’ for models with a single, infinitely-lived agent. In a typical overlapping-generations model the new young are born at the start of every period, and the old die at the end of every period, and the model has infinitely many periods; so there are infinitely many agents. On this view, the overlappinggenerations model is not a representative-agent model. I, however, regard it as one, because within any period one type of young agent and one type of old agent stand in for the enormous variety of people, and the same types are repeated period after period. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors FACTS AND ARTIFACTS 287 K.D.HOOVER 39 model permits better use of other information. Lars Peter Hansen and Sargent (1980, pp. 91, 92), for example, argue that the strength of their estimation method is that it accounts consistently for the interrelationships between constituent parts of the model—i.e. that is a general-equilibrium method. Calibrators respond, however, that it is precisely the importance of general equilibrium that supports their approach. Kydland and Prescott write: …it is in the stage of calibration where the power of the general equilibrium approach shows up most forcefully. The insistence on internal consistency implies that parsimoniously parameterized models of the household and business sector display rich dynamic behavior through the intertemporal substitution arising from capital accumulation and from other sources. (1991, p. 170) The trade-off between the gains and losses of the two methods is not clear cut. Lucas (1987, pp. 46, 47) and Prescott (1986, p. 15) argue that the strength of calibration is that it uses multiple sources of information, supporting the belief that it is structured around true invariants. This argument would appear to appeal to the respectable, albeit philosophically controversial view, that a theory is better supported when tested on information not used in its formulation (see Lipton 1991, ch. 8; Hoover 1994b). Unfortunately, it is not clear that calibration relies on independent information nor that it avoids estimation altogether. Parameters are sometimes chosen for calibrated business-cycle models because they mimic socalled ‘stylized facts’. That the models then faithfully reproduce such facts is not independent information. Other parameters are chosen from microeconomic studies. This introduces estimation through the back door, but without any but a subjective, aesthetic metric to judge model performance. Furthermore, since all new classical, equilibrium business-cycle models rely on the idealization of the representative agent, both calibrated and estimated versions share a common disability: using the representative-agent model in any form begs the question by assuming that aggregation does not fundamentally alter the structure of the aggregate model. Physics may provide a useful analogy. The laws that relate pressure, temperature, and volumes of gases are macrophysics. The ‘ideal-gas laws’ can be derived from a micromodel: gas molecules are assumed to be point masses, subject to conservation of momentum, with a distribution of velocities. An aggregate assumption is also needed: the probability of the gas molecules moving in any direction is taken to be equal. Direct estimation of the ideal gas laws shows that they tend to break down— and must be corrected with fudge factors—when pushed to extremes. For example, under high pressures or low temperatures the ideal laws must be corrected according to van der Waals’ equation. This phenomenal law, a law in macrophysics, is used to justify alterations of the micromodel: when pressures are high one must recognize that forces operate between individual molecules. The inference of the van der Waals’ force from the macrophysical behavior of gases has an analogue in the development of real-business-cycle models. Gary Hansen (1985), for example, introduces the microeconomic device of indivisible labor into a real-business cycle model, not from any direct reflection on the nature of labor markets at the level of the individual firm or © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 288 40 METHODOLOGY FACTS AND ARTIFACTS worker, but as an attempt to account for the macroeconomic failure of Kydland and Prescott’s (1982) model to satisfactorily reproduce the relative variabilities of hours and productivity in the aggregate data.20 Of course, direct estimation of Kydland and Prescott’s model rather than calibration may have pointed in the same direction.21 Despite examples of macro to micro inference analogous to the gas laws, Lucas’s (1980, p. 291) more typical view is that we must build our models up from the microeconomic to the macroeconomic. Unlike gases, human society does not comprise homogeneous molecules, but rational people, each choosing constantly. To understand (verstehen) their behavior, one must model the individual and his situation. This insight is clearly correct, it is not clear in the least that it is adequately captured in the heroic aggregation assumptions of the representativeagent model. The analogue for physics would be to model the behavior of gases at the macrophysical level, not as derived from the aggregation of molecules of randomly distributed momenta, but as a single molecule scaled up to observable volume—a thing corresponding to nothing ever known to nature.22 5. Calibration and macroeconomic practice The calibration methodology has both a wide following and a substantial opposition within the new classical school. I have attempted to give it a sympathetic reading—both in general and in its specific application to realbusiness-cycle models. I have concentrated on Kydland and Prescott, as its most prolific practitioners, and on Lucas, an articulate advocate. Although calibration is consistent with appealing accounts of the nature and role of models in science and economics, of their quanfication and idealization, its practical implementation in the service of real-business-cycle models with representative agents is less than compelling. Does the calibration methodology amount to a repudiation of econometric estimation altogether? Clearly not. At some level, econometrics still helps to supply the values of the parameters of the models. Beyond that, whatever has been said in favor of calibration methods to the contrary notwithstanding, the 20 Canova (1991b, p. 33) suggests that the particular covariance that Hansen’s modification of Kydland and Prescott’s model was meant to capture is an artifact of the HodrickPrescott filter, so that Hansen’s model may be a product of misdirected effort rather than a progressive adaptation. 21 This, rather than collegiality, may account for Kydland and Prescott’s (1982, p. 1369) tolerant remark about the future utility of Lars Peter Hansen and Sargent‘s (1980) econometric techniques as well as for Lucas’s (1987, p. 45) view that there is something to be learned from Altug’s estimations of the Kydland and Prescott model—a view expressed in the midst of arguing in favour of calibration. 22 A notable, non-new-classical attempt to derive macroeconomic behavior from microeconomic behavior with appropriate aggregation assumptions is Durlauf (1989). In a different, but related context, Stoker (1986) shows that demand systems fit the data only if distributional variables are included in the estimating equations. He takes this macroeconomic evidence as evidence for the failure of the microeconomic conditions of exact aggregation. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors FACTS AND ARTIFACTS K.D.HOOVER 289 41 misgivings of econometricians such as Sargent are genuine. The calibration methodology, to date, lacks any discipline as stern as that imposed by econometric methods. For Lucas (1980, p. 288) and Prescott (1983, p. 11), the discipline of the calibration method comes from the paucity of free parameters. But one should note that theory places only loose restrictions on the values of key parameters. In practice, they are actually pinned down from econometric estimation at the microeconomic level or accounting considerations. Thus, in some sense, the calibration method would appear to be a kind of indirect estimation. Thus, although as was pointed out earlier, it would be a mistake to treat calibration as simply an alternative form of estimation, it is easy to understand why some critics interpret it that way. Even were there less flexibility in the parameterizations, the properties ascribed to the underlying components of the idealized real-business-cycle models (the agents, their utility functions, production functions, and constraints) are not subject to as convincing cross checking as the analogous components in physical models usually are. The fudge factors that account for the discrepancies between the ideal model and the data look less like van der Waals’ equation than like special pleading. Above all, it is not clear on what standards competing, but contradictory, models are to be compared and adjudicated.23 Some such standards are essential if any objective progress is to be made in economics.24 ACKNOWLEDGEMENTS I thank Thomas Mayer, Kevin Salyer, Steven Sheffrin, Roy Epstein, Nancy Cartwright, Gregor Smith, Edward Prescrott, Adrian Pagan, and two anonymous referees for helpful comments on an earlier draft. The earliest version of this paper, entitled ‘Calibration versus Estimation: Standards of Empirical Assessment in the New Classical Macroeconomics’, was presented at the American Economic Association meetings in Washington, DC, December 1990. 23 Prescott (1983, p. 12) seems oddly, to claim that inability of a model to account for some real events is a positive virtue—in particular, that the inability of real-business-cycle models to account for the Great Depression is a point in their favour. He writes: ‘If any observation can be rationalized with some approach, then that approach is not scientific’. This seems to be a confused rendition of the respectable Popperian notion that a theory is more powerful the more things it rules out. But one must not mistake the power of a theory with its truth. Aside from issues of tractability, a theory that rationalizes only and exactly those events that actually occur, while ruling out exactly those events that do not occur is the perfect theory. In contrast, Prescott seems inadvertently to support the view that the more exceptions the better rule. 24 Watson (1993) develops a goodness-of-fit measure for calibrated models. It takes into account that, since idealization implies differences between model and reality that may be systematic, the errors-in-variables and errors-in-equations statistical models are probably not appropriate. Also see Gregory and Smith (1991, pp. 27–8), Canova (1991a), and Hoover (1994a). REFERENCES ADELMAN, I. and ADELMAN, F.L. (1959). ‘The Dynamic Properties of the Klein-Goldberger Model’, Econometrica, 27, 596–625. ALDRICH, J. (1989). ‘Autonomy’, Oxford Economic Papers, 41, 15–34. ALTUG, S. (1989). ‘Time-to-Build and Aggregate Fluctuations: Some New Evidence’, International Economic Review, 30, 889–920. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 290 42 METHODOLOGY FACTS AND ARTIFACTS BAN SAL, R. and GALLANT, R.A., H U S S EY, R., and TAUCH E N, G. (1991). ‘Nonparametric Structural Estimation of Models for High-Frequency Currency Market Data’, unpublished typescript. BLAUG, M. (1992). The Methodology of Economics: Or How Economists Explain, Cambridge University Press, Cambridge, 2nd ed. CANOVA, F. (1991a). ‘Sensitivity Analysis and Model Evalution in Simulated Dynamic General Equilibrium Economies’, unpublished typescript, Department of Economics, European University Institute, Florence, Italy. CANOVA, F. (1991b). ‘Detrending and Business Cycle Facts’, unpublished typescript, Department of Economics, European University Institute, Florence, Italy. CANOVA, F., FINN, M., and PAGAN, A.R. (1992). ‘Evaluating a Real Business Cycle Model’, unpublished typescript. COGLEY, T. and NASON, J.M. (1993). ‘Effects of the Hodrick-Prescott Filter on Trend and Difference Stationary Time Series: Implications for Business Cycle Research’, unpublished typescript. COOLEY, T.F. and HANSEN, G.D. (1989). ‘Welfare Consequences of Monetary and Fiscal Policy’, paper presented to the Federal Reserve Bank of San Francisco’s Fall Academic Conference, 1 December. COURNOT, A. [1838] (1927). Researches into the Mathematical Principles of the Theory of Wealth, Nathaniel T.Bacon (trans,), Macmillan, New York. DIAMOND, P.A. (1984). A Search-Equilibrium Approach to the Micro Foundations of Macroeconomics: The Wicksell Lectures, 1982, MIT Press, Cambridge, MA. DURLAUF, S.N. (1989). ‘Locally Interacting Systems, Coordination Failure, and the Behavior of Aggregate Activity’ , unpublished typescript. FRIEDMAN, M. (1953). ‘The Methodology of Positive Economics’, in Essays in Positive Economics, Chicago University Press, Chicago, IL. FRISCH, R. (1933). ‘Propagation Problems and Impulse Response Problems in Dynamic Economics’, in Economic Essays in Honour of Gustav Cassel: October 20th 1933, George Allen and Unwin, London. GREGORY, A.W. and SMITH, G.W. (1990a). ‘Calibration as Estimation’, Econometric Reviews, 9, 57–89. GREGORY, A.W. and SMITH, G.W. (1990b). ‘Calibration as Testing: Inference in Simulated Macroeconomic Models’ in Journal of Business and Economic Statistics, forthcoming. G REGORY, A.W. and S M ITH, G.W. (1991). ‘Calibration in Macroeconomics’, in G.S.Maddala and C.R.Rao (eds), Handbook of Statistics, 10: Econometics, forthcoming. HAAVELMO, T. (1944). ‘The Probability Approach in Econometrics’, Econometrica, 12 (supplement). HALL, R.E. (1977). ‘Investment, Interest Rates, and the Effects of Stabilization Policies’ Brookings Papers on Economic Activity, 6, 61–100. HANSEN, G.D. (1985). ‘Indivisible Labor and the Business Cycle’, Journal of Monetary Economics, 16, 309–28. HANSEN, G.D. and PRESCOTT, E.C. (1993). ‘Did Technology Shocks Cause the 1990– 1991 Recession’, American Economic Review, 83, 280–6. HANSEN, G.D. and SARGENT, T.J. (1988), ‘Straight Time and Overtime in Equilibrium’, Journal of Monetary Economics, 21, 281–308. HANSEN, L.P. and SARGENT, T.J. (1980). ‘Formulating and Estimating Dynamic Linear Rational Expectations Models’, in R.E.Lucas, Jr. and T.J.Sargent (eds), Rational Expectations and Econometric Practice, George Allen & Unwin, London, 1981. HARTLEY, J., SALYER, K., and SHEFFRIN, S. (1992). ‘Calibration and Real Business Cycle Models: Two Unorthodox Tests’, unpublished typescript. HOOVER, K.D. (1984). ‘Two Types of Monetarism’, Journal of Economic Literature, 22, 58–76. HOOVER, K.D. (1988). The New Classical Macroeconomics: A Skeptical Inquiry, Blackwell, Oxford. HOOVER, K.D. (1992). ‘The Rational Expectations Revolution: An Assessment’, The Cato Journal, 12, 81–96. HOOVER, K.D. (1994a). ‘Six Queries About Idealization in an Empirical Context’, Poznan Studies in the Philosphy of Science and the Humanities, 38, 43–53. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors FACTS AND ARTIFACTS K.D.HOOVER 291 43 HOOVER, K.D. (1994b). ‘In Defence of Datamining: Some Preliminary Throughts’, in K.D.Hoover and S.M.Sheffrin (eds), Monetarism and the Methodology of Econmics: Essays in Honor of Thomas Mayer, Elgar, Aldershot, forthcoming. KING, R.G. and PLOSSER, C.I. (1984). ‘Money, Credit and Prices in a Real Business Cycle’, American Economic Review, 74, 363–80. KING, R.G. and PLOSSER, C.I. (1989). ‘Real Business Cycle Models and the Test of the Adelmans’, paper presented to the Federal Reserve Bank of San Francisco’s Fall Academic Conference. 1 December 1989. KOOPMANS, T.C. (1947). ‘Measurement without Theory’, Review of Economics and Statistics, 29, 161–72. KYDLAND, F.E. (1984). ‘Labor-force Heterogeneity and the Business Cycle’ in K.Brunner and A.H.Meltzer (eds), Essays on Macroeconomic Implications of Financial and Labor Markets and Political Processes, Carnegie-Rochester Conference Series on Public Policy, 21. KYDLAND, F.E. and PRESCOTT, E.C. (1982). ‘Time to Build and Aggregate Fluctuations’, Econometrica, 50, 1345–70. KYDLAND, F.E. and PRESCOTT, E.C. (1988). ‘The Workweek of Capital and Its Cyclical Implication’ Journal of Monetary Economics, 21, 243–60. KYDLAND, F.E. and PRESCOTT, E.C. (1990). ‘Business Cycles: Real Facts and a Monetary Myth’, Federal Reserve Bank of Minneapolis Quartly Review, 14, 3–18. KYDLAND, F.E. and PRESCOTT, E.C. (1991). ‘The Econometrics of the Great Equlibrium Approach to Business Cycles’, Scandinavian Journal of Economics, 93, 161–78. LAKATOS, I. (1970). ‘Falsification and the Methodology of Scientific Research Programmes’, in I. Lakatos and A.Musgrave (eds), Criticism and the Growth of Knowledge, Cambridge University Press, Cambridge. LAKATOS, L (1978). ‘History of Science and Its Rational Reconstructions’ in J.Worrall and G. Currie (eds), The Methodology of Scientific Research Programmes: Philosphical Papers, I, Cambridge University Press, Cambridge. LAU, L.J. (1984). ‘Comments’, in H.E.Scarf and J.B.Shoven (eds), Applied General Equilibrium Analysis, Cambridge University Press, Cambridge. LIPTON. P. (1991). Inference to the Best Explanation, Routlege, London. LUCAS, R.E., JR. (1976). ‘Econometric Policy Evaluaton: A Critique’, in Studies in BusinessCycle Theory, Oxford, Blackwell, 1981. LUCAS, R.E., JR. (1977). ‘Understanding Business Cycles’, in Studies in Business-Cycle Theory, Blackwell, Oxford, 1981. LUCAS, R.E.. JR. (1980) ‘Methods and Problems in Business Cycle Theory’, in Studies in Business-Cycle Theory, Blackwell, Oxford, 1981. LUCAS, R.E. JR. (1986). ‘Adapttive Behavior and Economic Theory’, Journal of Business, 59, S401–26. LUCAS. R.E., JR. (1987), Models of Business Cycles, Blackwell, Oxford. LUCAS, R.E., JR. and SARGENT, T.J.C. (eds) (1981). Rational Expectations and Econometric Practice, George Allen and Unwin, London. MACKINNON.J. (1984). ‘Comments’, in H.E.Scarf and J.B.Shoven (eds), Applied General Equilibrium Analysis , Cambridge University Press, Cambridge. MAN SU R, A. and WALLEY, J. (1984). ‘Numerical Specification of Applied Generalequilibrium models: Estimation, Calibration, and Data’, in H.E.Scarf and J.B.Shoven (eds.), Applied General Equilibrium Analysis, Cambridge University Press, Cambridge. MANUELLI, R. and SARGENT.T.J. (1988). ‘Models of Business Cycles: A Review Essay’ Journal of Monetary Economics, 22, 523–42. MAYER, T. (1960). ‘Plant and Equipment Lead Times’ Journal of Business, 33, 127–32. MCCALLUM, B.T. (1986). ‘On “Real” and “Sticky-Price” Theories of the Business Cycle’, Journal of Money, Credit and Banking, 18, 397–414. MCCALLUM, B.T. (1989). ‘Real Business Cycle Models’, in R.J.Barro (ed.), Modern Business Cycle Theory, Blackwell, Oxford. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors Part V The critique of calibration methods © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors CHAPTER 16 © 1991 American Statistical Association 295 Journal of Business & Economic Statistics, July 1991, Vol. 9, No. 3 Calibration as Testing: Inference in Simulated Macroeconomic Models Allan W.Gregory and Gregor W.Smith Department of Economics, Queen’s University, Kingston, Ontario, K7L 3N6, Canada A stochastic macroeconomic model with no free parameters can be tested by comparing its features, such as moments, with those of data. Repeated simulation allows exact tests and gives the distribution of the sample moment under the null hypothesis that the model is true. We calculate the size of tests of the model studied by Mehra and Prescott. The approximate size of their test (which seeks to match model-generated, mean, riskfree interest rates and equity premia with historical values) is 0 although alternate, empirical representations of this model economy or alternate moment-matching tests yield large probabilities of Type I error. KEY WORDS: Equity premium; Monte Carlo; Simulation; Type I error. Calibration in macroeconomics is concerned primarily with testing a model by comparing population moments (or perhaps some other population measure) to historical sample moments of actual data. If the correspondence between some aspect of the model and the historical record is deemed to be reasonably close, then the model is viewed as satisfactory. If the distance between population and historical moments is viewed as too great, then the model is rejected, as in the widely cited equity-premium puzzle of Mehra and Prescott (1985). A drawback to the procedure as implemented in the literature is that no metric is supplied by which closeness can be judged. This leads to tests with unknown acceptance and rejection regions. This article provides a simple way to judge the degree of correspondence between the population moments of a simulated macroeconomic model and observed sample moments and develops a framework for readily calculating the size (probability of Type I error) of calibration tests. We apply this method to the well-known equitypremium case. This article is not concerned with a “solution” to the equity-premium puzzle. Rather it evaluates the probability of falsely rejecting a true macroeconomic model with calibration methods. One finding is that the size of the test considered by Mehra and Prescott (which seeks to match mean risk-free interest rates and equity premia) is 0, so the model with their parameter settings is unlikely to have generated the observed historical moments. Some alternate versions of the consumption-based asset-pricing model or alternate moment-matching tests yield large probabilities of Type I error. Section 1 characterizes calibration as testing. A simple formalization of calibration as Monte Carlo testing allows exact inference. Section 2 contains an application to the test conducted by Mehra and Prescott (1985). Section 3 concludes. 1. CALIBRATION AS TESTING Calibration in macroeconomics has focused on comparing observed historical moments with population moments from a fully parameterized simulation model—that is, one with no free parameters. One might elect to simulate a model because of an analytical intractability or because a forcing variable is unobservable. In macroeconomics, examples of unobservable forcing variables include productivity shocks in business-cycle models or consumption measurement errors in asset-pricing models. Consider a population moment θ, which is restricted by theory, with corresponding historical sample moment for a sample of size T. Call the moment estimator . Assume that is consistent for θ. The population moment is a number, the sample moment is the realization of a random variable (an estimate), and the estimator is a random variable. The calibration tests applied in the literature compare θ and and reject the model if θ is not sufficiently close to . In some calibration studies, attempts are made to exactly match the population moment to the sample moment (there must be some leeway in parameter choice to make this attempt nontrivial). Such matching imposes unusual test requirements because θ and can differ even when the model is true due to sampling variability in . Moreover, judging closeness involves the sampling distribution of the estimator . Standard hypothesis testing procedures may be unavailable because the exact or even asymptotic distribution of the estimator is unknown. One prominent advantage in the calibration of macroeconomic models that has not been exploited fully is that the complete datagenerating process is specified. Thus the 297 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 296 298 CELEBRATION Journal of Business & Economic Statistics, July 1991 sampling variability of the simulated moment can be used to evaluate the distance between θ and Call N the number of simulated observations. We construct tests by repeatedly simulating the fully parameterized model (or artificial economy) and calculating the proportion of times lies in a set ⍜(bounded by ). Current calibration studies do simulate repeatedly, but they typically average over a small number of replications and then quote the averages of various properties. Our approach involves treating the historical as a critical value to give the moment probability (prob) value (i.e., marginal significance level) or size of the test implicit in comparing moments. If simulations use the same number of observations as are used in calculating the historical sample moment so that N=T, then the test will be exact. Moreover, confidence intervals can be constructed for (as opposed to the population moment θ ) to determine whether lies within the interval. Barnard (1963) suggested basing exact inference on repeated generation of artificial series. Monte Carlo testing was refined by Hope (1968) and Marriott (1979) and applied in economics by Theil, Shonkwiler, and Taylor (1985) and Theil and Shonkwiler (1986). Simulation methods also can be used for estimation, as described for macroeconomic models by Lee and Ingram (1991) and Gregory and Smith (1990). Suppose one wishes to find the probability with which a model with population moment θ generates a sample moment Our suggested procedure is to simulate N observations from the economy, with N=T, R times and calculate at each replication. The proportion of replications with which the simulated moment exceeds the corresponding sample moment is the (one-sided) prob value associated with the historical moment (1) in which the indicator function I is defined to be equal to 1 if its argument is positive and 0 otherwise. Simulations involve a finite number of replications, so measures of test size or confidence intervals are themselves random variables. Standard errors for test size may be calculated from the binomial distribution or, if R is large, the normal distribution. Although we have described treating the historical moments as critical values, one also could select a test size and estimate the corresponding critical value (i.e., the quantile) from the order statistics of the Monte Carlo sample of moments. In some cases one might be interested in the entire probability density of the sample moment under the null hypothesis that the model is true. Although it is simpler to work with the empirical density function, tests also could be based on the density estimated nonparametrically from ], and approximate simulations [denoted by measures of size could be calculated. There are many estimators available (see Silverman 1986 or Tapia and Thompson 1978), but if we assume that the density is absolutely continuous, then kernel estimators provide an easy way to smooth the empirical density. One advantage in applying kernel estimators is that they are available in an IMSL subroutine. In the application in Section 2, we calculate prob values in two ways. The first way is the simplest and uses the empirical distribution as in Equation (1). The second way estimates the density nonparametrically with a quartic kernel and, following Silverman (1986), a robust, variable window width given by .698 min [s, IQR/ 1.34], in which s is the empirical standard deviation and IQR the interquartile range. Under weak conditions the empirical distribution function and the nonparametric density consistently estimate the true cumulative distribution function and density, respectively. In the simulations in Section 2, the results from the two methods usually are identical, which suggests using the simpler calculation in Equation (1). As N becomes large, sample moments converge to population moments by ergodicity. For fixed N, sample moments have distributions; as the number of replications R increases, the estimated distribution function converges to the true finite-sample distribution. One could examine the sensitivity of findings to alternative windows, kernels, and numbers of replications because the speed of convergence may be slow. For example, many replications may be necessary for accurate estimates of tail-area probabilities (i.e., test size), where by definition events are rare in a sampling experiment. Finally, one also could study tests based on matching properties other than moments, by analogy with the estimation results of Smith (1989), who considered matching properties of misspecified models or linear regressions in simulated and historical data. Naylor and Finger (1971, p. 159) listed other validation criteria, such as matching properties of turning points. 2. APPLICATION: TYPE I ERROR IN THE EQUITY-PREMIUM PUZZLE As an example, consider the exchange economy described by Mehra and Prescott (1985). A © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors CALIBRATION AS TESTING 297 Gregory and Smith: Calibration as Testing 299 representative consumer has preferences characterized by the utility functional (2) in which Et denotes the expectation conditional on information at time t, β 僆 (0,1) is a discount factor, u(c) =c1-α/(1-α) [log(c) if α=1], and α is a positive, constant coefficient of relative risk aversion. The consumer’s nonstorable endowment yt evolves exogenously according to (5b) Thus if the current state is (yt, li), then the prices (relative to one unit of the commodity at time t) of the two assets are (6a) and (3) (6b) in which the growth rate, xt, follows a process that is Markov on a discrete state space Λ={λ1, λ2, …, λJ}. This process is stationary and ergodic, with transition matrix , (4) The equilibrium or unconditional probabilities are given by φi=Pr[xt=λi] for all t. An equilibrium in this economy is characterized by a set of prices for which consumption equals the endowment at all times and in all states. Relative prices are calculated by equating them to marginal rates of substitution. The price of a one-period, risk-free discount bond that provides one unit of the endowment at time t+1 is given by (5a) and the price of an equity claim to the endowment stream is given by The challenge posed of this model by Mehra and Prescott has been the following: With a risk-aversion parameter α between 0 and 10, a discount factor β between 0 and 1, and the Markov per capita consumption growth-rate process matching the sample mean, variance, and first-order autocorrelation of the U.S. series for 1889– 1978, can the model (with two or three states) generate a population mean risk-free rate of return and mean equity premium that match the annual U.S. sample values (.8% and 6.2%, respectively) of these measures? Mehra and Prescott compared the 6.2% average equity premium from the full sample (90 observations) to an average premium from their model of at most .35%. These two values were judged not to be close, and hence there is a puzzle. One could try to gauge closeness with a standard asymptotic hypothesis test. For example, the test statistic implicit in Mehra and Prescott’s (1985, p. 156) quotation of the estimated average equity premium and its standard error is Table 1. Asset-Pricing Models and Joint Tests NOTE: The prob value is the marginal significance level of the joint one-sided test that the population mean equity premium is at least 6.2% and the population mean risk-free rate is less than 4%. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 298 300 CELEBRATION Journal of Business & Economic Statistics, July 1991 Table 2. Sample and Population Moments: Case 1, Symmetric (Mehra-Prescott) NOTE: The symbols µ. σ, and ρ denote mean, variance, and first-order autocorrelation. Sample moments in the third column are based on the full sample. The sample autocorrelations for the riskfree and equity-premium rates were not given in Mehra and Prescottís table 1 (1985, p. 147). Pop. denotes population moments from the model. Emp denotes measures constructed from the empirical density function and Np those constructed from the nonparametrically estimated density function, each based on R=1,000 replications. Confidence intervals run from the .025 to als run from the .025 to .975 quantiles. a Prob values for b Prob values for asymptotically. In this test one could correct the standard error of the average equity premium to allow for dependence and heterogeneity with the methods of Newey and West (1987) or Andrews (1988). The heterogeneity does not matter in a test with no regressors (as in a comparison of means), but the serial correlation does. Cecchetti, Lam, and Mark (1990) constructed an asymptotic test statistic for matching a moment: where is W= the Newey-West estimate of the variance of the sample moment. To allow for uncertainty about parameters that are estimated, they assumed that the two sources of uncertainty are independent and amended the statistic to W= where is the variance due to parameter uncertainty. Kocherlakota (1990) conducted a Monte Carlo experiment of tests of consumption-based asset-pricing models that use asymptotic 5% critical values and examined the small-sample size of such tests with simulated data. He found that size most closely approximates asymptotic size when parameters are set rather than estimated. Finding a consistent estimate even of the asymptotic standard error for many estimated moments can be very difficult, however. Moreover, some of the historical samples considered by Mehra and Prescott have as few as 10 observations and hence should not be studied with asymptotic methods. On the other hand, the procedure outlined in Section 1 is appropriate for inference in such sample sizes using the sampling Table 3. Sample and Population Moments: Case 2, Crash (Rietz) NOTE: See note to Table 2. a Prob values for b Prob values for © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors CALIBRATION AS TESTING 299 Gregory and Smith: Calibration as Testing 301 Table 4. Sample and Population Moments: Case 3, Asymmetric NOTE. See note to Table 2. a Prob values for b Prob values for variability inherent in fully parameterized models. To investigate the size of the momentmatching exercise considered by Mehra and Prescott, suppose that their model is true. What is the probability of observing an equity premium of at least 6.2%? To answer this question, we simulate the fully parameterized model, estimate the probability density functions for the random variables that it restricts, and measure tail-area probabilities. We follow the same procedure for four different representations of the transition density for the Markov chain and of risk aversion. These are shown in Table 1. The first representation is that considered by Mehra and Prescott. The second representation is that used by Rietz (1988), in which there is a small probability of a “crash” in growth rates; the source is Rietz’s example 1, table 3, row 1. The third representation involves an asymmetry in growth transitions, and the fourth representation involves a degree of risk aversion greater than that considered by Mehra and Prescott. The explicit joint test considered by Mehra and Prescott (1985, p. 154) is to generate a population equity premium of at least 6.2% and a population risk-free rate less than 4%. For each model we estimate the prob value associated with this joint test. These measures are based on bivariate empirical distributions with R= 1,000 replications and on N=90 observations, just as in the annual sample. The results are shown in Table 1. For the Mehra-Prescott model of case 1, the prob value is 0. The other models generate the historical sample-mean rates of return with positive probabilities. The prob values are .59 for case 2, .14 for case 3, and .95 for case 4. Tables 2–5 give the sample mean, variance, and first-order autocorrelation of consumption growth, the risk-free rate, and the equity premium for the annual data used by Mehra and Prescott and the corresponding population values for the four simulation models described in Table 1. The autocorrelations of the risk-free rate and consumption growth are equal as an analytical property of the model. For each Table 5. Sample and Population Moments: Case 4, Extreme Risk Aversion NOTE: See note to Table 2. a Prob values for b Prob values for © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 300 302 CELEBRATION Journal of Business & Economic Statistics, July 1991 Figure 1. The Nonparametrically Estimated Density Function (based on 1,000 replications) of the Mean Equity Premium in Simulated Samples of 90 Observations, From the Model Economy (4), With Extreme Risk Aversion. Parameter settings are given in Table 1. model we construct 95% confidence intervals for a sample size of 90 observations. We also estimate the densities nonparametrically and calculate corresponding approximate intervals. Tables 2–5 also give size results for some univariate, one-sided tests of each model’s ability to match individual moments. For example, Table 2 shows that for one of the Mehra-Prescott parameter settings (given in Table 1) the equity premium has a mean of .20. The estimated probability of observing a premium greater than 6.2% is 0; the 95% confidence interval is (-.6, 1.0) at N=90 observations. One can also see that the puzzle extends to second moments, because the variances of historical riskfree rates and equity premia lie outside the confidence intervals generated by the theoretical model. Mehra and Prescott (1985, p. 146) did note that their model was poorly suited for assetprice volatility issues. The confidence intervals for autocorrelation coefficients are generally wide and would rationalize a wide range of observed sample values. The Rietz model (2) in Table 3 is consistent with the sample mean equity premium of 6.18% and, in fact, would be consistent with an even larger premium; the 95% confidence interval is (5.1, 7.5) at N=90 observations and the prob value (from the simulated empirical density) is .59. Like model 1, it fails to match the variances of the riskfree rate and equity premium. The model with asymmetric growth (3) in Table 4 generates larger variances for the two returns but still cannot match the variance of the equity-premium rate. Model 3 is consistent with the sample mean-equity premium in that it generates the sample value with a prob value of .10, but it also gives a consumption growth variance that is too large relative to the sample variance. In Table 5, model 4, in which α=25, gives 95% confidence intervals that include the historical values for all moments except the variance of the risk-free rate. Although models 1 and 2 underpredict the variance of the risk-free rate, models 3 and 4 overpredict it. Finally, for model 4, Figure 1 illustrates the nonparametrically estimated density function of the mean equity premium from which the confidence interval in the last column of Table 5 is taken. In Tables 2–5 the confidence intervals and size estimates from the empirical densities and the nonparametrically estimated densities are virtually identical. We also have used the same method with a sample size of N=10 for comparison with sample values reported by Mehra and Prescott (1985, table 1) for 1959–1968. Naturally the intervals for the smaller simulated sample size are wider. We also have investigated the effects of increasing the number of replications to R=2,000. This increase has no appreciable effect on the confidence intervals. Although faster convergence in R might be possible with the use of control or antithetic variates to reduce sampling variability in estimated test sizes and confidence intervals, such methods do not seem vital in this application. In constructing the tests in the tables, we do not rule out applying other criteria to evaluate versions of this asset-pricing model. For example, one might discard the fourth case because of its large risk aversion or the second case because its growthtransition matrix includes elements never observed (see Mehra and Prescott 1988). We are agnostic on these points. The proposed procedure simply allows one to gauge the sampling variability of the moment estimator when the model is assumed to be true and thus permits formal comparisons of historical and population (model) moments. Kwan (1990) and Watson (1990) have proposed alternative measures of the goodness of fit of calibrated models (or closeness of moments) in which the parameterized economic model is not the null hypothesis. 3. CONCLUSION Calibration in macroeconomics involves simulating dynamic models and comparing them to historical data. This type of comparison can be seen as a test of the model. A natural issue in the use of this method concerns the distribution of the sample moment under the null hypothesis that the model is true. For fully parameterized model © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors CALIBRATION AS TESTING 301 Gregory and Smith: Calibration as Testing 299 economies, one can use repeated simulation to estimate the probability of falsely rejecting a true economic model (Type I error) and to construct confidence intervals using comparisons of moments. The empirical frequency estimator works well in both cases and is very simple to apply. As an application we consider the assetpricing model and parameter settings of Mehra and Prescott, for which the test of mean rates of return has no probability of Type I error; their rejection of the model appears to be statistically sound. Our results can be used to find the sizes of other moment-matching tests of this model. If other parameter settings are admitted, then the model will generate an equity premium of 6.2% more than 5% of the time. The method can be applied directly to extensions of the asset-pricing economy discussed previously (see, for example, Abel 1990; Benninga and Protopapadakis 1990; Cecchetti et al. 1990; Labadie 1989) and to business-cycle models, as Devereux, Gregory, and Smith (1990) demonstrated. ACKNOWLEDGMENTS We thank the Social Sciences and Humanities Research Council of Canada for financial support. We thank David Backus, Jean-Marie Dufour, James MacKinnon, Adrian Pagan, Simon Power, Tom Rietz, Jeremy Rudin, the editor, an associate editor, a referee, and numerous seminar participants for helpful comments. [Received April 1990. Revised January 1991.] REFERENCES Abel, A.B. (1990), “Asset Prices Under Habit Formation and Catching up With the Joneses,” American Economic Review, 80, 38–42. Andrews, D.W.K. (1988), “Heteroskedasticity and Autocorrelation Consistent Covariance Matrix Estimation,” Cowles Foundation Discussion Paper 877R, Yale University, Dept. of Economics. Barnard, G.A. (1963), Discussion of “The Spectral Analysis of Point Process,” by M.S.Bartlett, Journal of the Royal Statistical Society, Ser. B, 25, 294. Benninga, S., and Protopapadakis, A. (1990), “Leverage, Time Preference, and the ‘Equity Premium Puzzle,’” Journal of Monetary Economics, 25, 49–58. Cecchetti, S.G., Lam, P., and Mark, N.C. (1990), “The Equity Premium and the Risk Free Rate: Matching the Moments,” mimeo, Ohio State University, Dept. of Economics. Devereux, M.B., Gregory, A.W., and Smith, G.W. (1990), “Realistic Cross-Country Consumption Correlations in a Two-County Equilibrium Business Cycle Model,” IER Discussion Paper 774, Queen’s University, Dept. of Economics. Gregory, A.W., and Smith, G.W. (1990), “Calibration as Estimation,” Econometric Reviews, 9, 57–89. Hope, A.C.A. (1968), “A Simplified Monte Carlo Significance Test Procedure,” Journal of the Royal Statistical Society, Ser. B, 30, 582–598. Kocherlakota, N.R. (1990), “On Tests of Representative Consumer Asset Pricing Models,” Journal of Monetary Economics, 26, 285–304. Kwan, Y.-K. (1990), “Bayesian Model Calibration With An Application to a Non-Linear Rational Expectations Two-Country Model,” mimeo. University of Chicago, Graduate School of Business. Labadie, P. (1989), “Stochastic Inflation and the Equity Premium,” Journal of Monetary Economics, 24, 277–298. Lee, B.-S., and Ingram, B.F. (1991), “Simulation Estimation of Time Series Models,” Journal of Econometrics, 47, 197–205. Marriott, F.H.C. (1979), “Barnard’s Monte Carlo Tests: How Many Simulations?” Applied Statistics, 28, 75–77. Mehra, R., and Prescott, E.C. (1985), “The Equity Premium: A Puzzle,” Journal of Monetary Economics, 15, 145–161. ——(1988), “The Equity Risk Premium: A Solution?” Journal of Monetary Economics, 22, 133–136. Naylor, T.H., and Finger, J.M. (1971), “Validation,” in Computer Simulation Experiments With Models of Economic Systems, ed. T.H.Naylor, New York: John Wiley, pp. 153–164. Newey, W.K., and West, K.D. (1987), “A Simple, Positive Semidefinite, Heteroskedasticity and Autocorrelation Consistent Covariance Matrix,” Econometrica, 55, 703–708. Rietz, T. (1988), “The Equity Risk Premium: A Solution,” Journal of Monetary Economics, 22, 117– 131. Silverman, B.W. (1986), Density Estimation for Statistics and Data Analysis, London: Chapman & Hall. Smith, A. (1989), “Estimation of Dynamic Economic Models by Simulation: A Comparison of Two Approaches,” mimeo. Duke University, Dept. of Economics. Tapia, R.A., and Thompson, J.R. (1978), Nonparameiric Probability Density Estimation, Baltimore: The Johns Hopkins University Press. Theil, H., and Shonkwiler, J.S., (1986), “Monte Carlo Tests of Autocorrelation,” Economics Letters, 20, 157–160. Theil, H., Shonkwiler, J.S., and Taylor, T.G. (1985). “A Monte Carlo Test of Slutsky Symmetry,” Economics Letters, 19, 331–332. Watson, M.W. (1990), “Measures of Fit for Calibrated Models,” mimeo, Northwestern University, Dept. of Economics. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 302 CHAPTER 17 Measures of Fit for Calibrated Models Mark W.Watson Northwestern University and Federal Reserve Bank of Chicago This paper suggests a new procedure for evaluating the fit of a dynamic structural economic model. The procedure begins by augmenting the variables in the model with just enough stochastic error so that the model can exactly match the second moments of the actual data. Measures of fit for the model can then be constructed on the basis of the size of this error. The procedure is applied to a standard real business cycle model. Over the business cycle frequencies, the model must be augmented with a substantial error to match data for the postwar U.S. economy. Lower bounds on the variance of the error range from 40 percent to 60 percent of the variance in the actual data. I. Introduction Economists have long debated appropriate methods for assessing the empirical relevance of economic models. The standard econometric approach can be traced back to Haavelmo (1944), who argued that an economic model should be embedded within a complete probability model and analyzed using statistical methods designed for conducting inference about unknown probability distributions. In the modern literature, this approach is clearly exemplified in work such as that of L.Hansen and Sargent (1980) or McFadden (1981). However, many economic models do not provide a realistic and complete This paper has benefited from constructive comments by many seminar participants; in particular I thank John Cochrane, Marty Eichenbaum, Jon Faust, Lars Hansen, Robert Hodrick, Robert King, and Robert Lucas. Two referees also provided valuable constructive criticism and suggestions. The first draft of this paper was written while I was visiting the University of Chicago, whose hospitality is gratefully acknowledged. This research was supported by the National Science Foundation through grants SES-89–10601 and SES-9122463. [Journal of Political Economy, 1993, vol. 101, no. 6] © 1993 by The University of Chicago. All rights reserved. 0022–3808/93/0106–0007$01.50 1011 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors MEASURES OF FIT 1012 303 JOURNAL OF POLITICAL ECONOMY probability structure for the variables under consideration. To analyze these models using standard econometric methods, they must first be augmented with additional random components. Inferences drawn from these expanded models are meaningful only to the extent that the additional random components do not mask or change the salient features of the original economic models. Another approach, markedly different from the standard econometric approach, has become increasingly popular for evaluating dynamic macroeconomic models. This approach is clearly articulated in the work of Kydland and Prescott (1982) and Prescott (1986). In a general sense, the approach asks whether data from a real economy share certain characteristics with data generated by the artificial economy described by an economic model. There is no claim that the model explains all the characteristics of the actual data, nor is there any attempt to augment the model with additional random components to more accurately describe the data. On the one hand, the results from this approach are easier to interpret than the results from the standard econometric approach since the economic model is not complicated by additional random elements added solely for statistical convenience. On the other hand, since the economic model does not provide a complete probability structure, inference procedures lack statistical foundations and are necessarily ad hoc. For example, a researcher may determine that a model fits the data well because it implies moments for the variables under study that are “close” to the moments of the actual data, even though the metric used to determine the distance between the moments is left unspecified. This paper is an attempt to put the latter approach on a less ad hoc foundation by developing goodness-of-fit measures for the class of dynamic econometric models whose endogenous variables follow covariance stationary processes. It is not assumed that the model accurately describes data from the actual economy; the economic model is not a null hypothesis in the statistical sense. Rather, the economic model is viewed as an approximation to the stochastic processes generating the actual data, and goodness-of-fit measures are proposed to measure the quality of this approximation. A standard device—stochastic error—is used to motivate the goodness-of-fit measures. These measures answer the question, How much random error would have to be added to the data generated by the model so that the autocovariances implied by the model+error match the autocovariances of the observed data? The error represents the degree of abstraction of the model from the data. Since the error cannot be attributed to a data collection procedure or to a forecasting procedure, for instance, it is difficult a priori to say much about its properties. In particular, its covariance with the observed data © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 304 CELEBRATION MEASURES OF FIT 1013 cannot be specified by a priori reasoning. Rather than make a specific assumption about the error’s covariance properties, I construct a representation that minimizes the contribution of the error in the complete model. Thus, in this sense, the error process is chosen to make the model as close to the data as possible. Many of the ideas in this paper are close to, and were motivated by, ideas in Altug (1989) and Sargent (1989). Altug (1989) showed how a one-shock real business cycle model could be analyzed using standard dynamic econometric methods, after first augmenting each variable in the model with an idiosyncratic error. This produces a restricted version of the dynamic factor analysis or unobserved index models developed by Sargent and Sims (1977) and Geweke (1977). Sargent (1989) discusses two models of measurement error: in the first the measurement error is uncorrelated with the data generated by the model, and in the second the measurement error is uncorrelated with the sample data (see also G.Hansen and Sargent 1988). While similar in spirit, the approach taken in this paper differs from that of Altug and Sargent in two important ways. First, in this paper, the error process is not assumed to be uncorrelated with the model’s artificial data or with the actual data. Rather, the correlation properties of the error process are determined by the requirement that the variance of the error is as small as possible. Second, the joint data-error process is introduced to motivate goodness-of-fit measures; it is not introduced to describe a statistical model that can be used to test statistical hypotheses, at least in the standard sense. Rather, the analysis in this paper is similar to the analysis in Campbell and Shiller (1988), Durlauf and Hall (1989), Hansen and Jagannathan (1991), and Cochrane (1992). Each of these papers uses a different approach to judge the goodness of fit of an economic model by calculating a value or an upper bound on the variance of an unobserved “noise” or a “marginal rate of substitution” or a “discount factor” in observed data. The minimum approximation error representation developed in this paper motivates two sets of statistics that can be used to evaluate the goodness of fit of the economic model. First, like the variance of the error in a regression model, the variance of the approximation error can be used to form an R2 measure for each variable in the model. This provides an overall measure of fit. Moreover, spectral methods can be used to calculate this R2 measure for each frequency so that the fit can be calculated over the “business cycle,” “growth,” or other specific frequency bands. A second set of statistics can be constructed by using the minimum error representation to form fitted values of the variables in the economic model. These fitted values show how well the model explains specific historical © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors MEASURES OF FIT 1014 305 JOURNAL OF POLITICAL ECONOMY episodes; for example, can a real business cycle model simultaneously explain the growth in the United States during the 1960s and the 1981– 82 recession? The paper is organized as follows. Section II develops the minimum approximation error representation and goodness-of-fit measures. Section III calculates these goodness-of-fit statistics for a standard real business cycle model using postwar U.S. macroeconomic data on output, consumption, investment, and employment. Section IV concludes the paper by providing a brief discussion of some tangential issues that arise from the analysis. II. Measures of Fit Consider an economic model that describes the evolution of an n×1 vector of variables xt. Assume that the variables have been transformed, say by first-differencing or forming ratios, so that xt is covariance stationary. As a notational device, it is useful to introduce the au toco variance generating function (ACGF) of xt, Ax(z). This function completely summarizes the unconditional second-moment properties of the process. In what follows, “economic model” and “Ax(z)” will be used interchangeably; that is, the analysis considers only the unconditional second-moment implications of the model. Nonlinearities and variation in conditional second and higher moments are ignored to help keep the problem tractable. The analysis will also ignore the unconditional first moments of xt; modifying the measures of fit for differences in the means of the variables is straightforward. The empirical counterparts of xt are denoted yt . These variables differ from xt in an important way. The variables making up xt correspond to the variables appearing in the theorist’s simplification of reality; in a macroeconomic model they are variables such as “output,” “money,” and the “interest rate.” The variables making up yt are functions of raw data collected in a real economy; they are variables such as “per capita gross domestic product in the United States in 1987 dollars” or “U.S. M2” or “the yield on 3-month U.S. Treasury bills.” The question of interest is whether the model generates data with characteristics similar to those of the data generated by the real economy. Below, goodness-of-fit measures are proposed to help answer this question. Before I introduce these new measures, it is useful to review standard statistical goodness-of-fit measures to highlight their deficiencies for answering the question at hand. Standard statistical goodness-of-fit measures use the size of sampling © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 306 CELEBRATION MEASURES OF FIT 1015 error to judge the coherence of the model with the data. They are based on the following: First, Ay(z), the population ACGF for yt, is unknown but can be estimated from sample data. Discrepancies between the estimator and arise solely from sampling error in and the likely size of the error can be deduced from the stochastic process that generated the sample. Now, if Ay(z)=Ax(z), sampling error also accounts for the differences and Ax(z). Standard goodness-of-fit measures show how between likely it is that Ay(z)=Ax(z), on the basis of the probability that differences and Ax(z) arise solely from sampling error. If the differences between and Ax(z) are so large as to be unlikely, standard measures of between fit suggest that the model fits the data poorly, and vice versa if the differences and Ax(z) are not so large as to be unlikely. The key point is between and Ax(z) are judged by how informative that the differences between the sample is about the population moments of yt. This is a sensible procedure for judging the coherence of a null hypothesis, Ay(z)=Ax(z), with the data. It is arguably less sensible when this null hypothesis is known to be false. Rather than rely on sampling error, the measures of fit that are developed here are based on the size of the stochastic error required to reconcile the autocovariances of xt with those of yt. In particular, let u, denote an n×1 error vector; then the importance of a difference between Ax(z) and Ay(z) will be determined by asking, How much error would have to be added to xt so that the autocovariances of xt+ut are equal to the autocovariances of yt? If the variance of the required error is large, then the discrepancy between Ax(z) and Ay(z) is large, and conversely if the variance of ut is small. The vector ut is the approximation error in the economic model interpreted as a stochastic process. It captures the second-moment characteristics of the observed data that are not captured by the model. Loosely speaking, it is analogous to the error term in a regression in which the set of regressors is interpreted as the economic model. The economic model might be deemed a good approximation to the data if the error term is small (i.e., the R2 of the regression is large) and might be deemed a poor approximation if the error term is large (i.e., the R2 of the regression is small). To be more precise, assume that xt and yt are jointly covariance stationary, and define the error ut by the equation (1) so that (2) where Au(z) is the ACGF of ut, Axy(z) is the cross ACGF between xt and yt, and so forth. From the right-hand side of (2), three terms are needed to © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors MEASURES OF FIT 1016 307 JOURNAL OF POLITICAL ECONOMY calculate Au(z). The first, Ay(z), can be consistently estimated from sample data; the second, Ax(z), is completely determined by the model; but the third, Axy(z), is not determined by the model, and it cannot be estimated from the data (since this would require a sample drawn from the joint (xt, yt) process). To proceed, an assumption is necessary. A common assumption used in econometric analysis is that Axy(z) =Ax(z) so that xt and ut are uncorrelated at all leads and lags. Equation (1) can then be interpreted as the dynamic analogue of the classical errors-in-variables model. Sargent (1989) discusses this assumption and an alternative assumption, Axy(z)=Ay(z). He points out that under this latter assumption, ut can be interpreted as signal extraction error, with yt an optimal estimate of the unobserved “signal” xt.1 In many applications, these covariance restrictions follow from the way the data were collected or the way expectations are formed. For example, if x t represented the true value of the U.S. unemployment rate and yt the value published by the U.S. Department of Labor, then yt would differ from xt because of the sampling error inherent in the monthly Current Population Survey from which yt is derived. The sample design underlying the survey implies that the error term, ut, is statistically independent of xt. Similarly, if yt denoted a rational expectation of xt, then the error would be uncorrelated with yt. Neither of these assumptions seems appropriate in the present context. The error is not the result of imprecise measurement. It is not a forecast or signal extraction error. Rather, it represents approximation or abstraction error in the economic model. Any restriction used to identify Axy(z), and hence Au(z), is arbitrary.2 It is possible, however, to calculate a lower bound for the variance of ut without imposing any restrictions on Axy(z). When this lower bound on the variance of ut is large, then under any assumption on Axy(z), the model fits the data poorly. If the lower bound on the variance of ut is small, then there are possible assumptions about Axy(z) that imply that the model fits the data well. Thus the bound is potentially useful for rejecting models 1 The reader familiar with work on data revisions will recognize these two sets of assumptions as the ones underlying the “news” and “noise” models of Mankiw, Runkle, and Shapiro (1984) and Mankiw and Shapiro (1986). 2 It is interesting to note that it is possible to determine whether the dynamic errors-in-variables model or the signal extraction error model is consistent with the model and the data. The dynamic errors-in-variables model implies that Ay(z)-Ax(z)≥0 for so that the spectrum of yt lies everywhere above the spectrum of xt; the signal extraction error model implies the converse. If the spectrum of x, lies anywhere above the spectrum of yt, the errors-in-variables model is inappropriate; if the spectrum of yt lies anywhere above the spectrum of xt, the signal extraction model is inappropriate. If the spectra of xt and yt cross, neither model is appropriate. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 308 CELEBRATION MEASURES OF FIT 1017 on the basis of their empirical fit. Needless to say, models that appear to fit the data well using this bound require further scrutiny. The bound is calculated by choosing Axy(z) to minimize the variance of ut subject to the constraint that the implied joint ACGF for xt and yt is positive semidefinite. Equivalently, since the spectrum is proportional to the ACGF evaluated at z=e-iω, the cross spectrum between xt and yt, (2π)1 Axy(e-iω), must be chosen so that the spectral density matrix of is positive semidefinite at all frequencies. Since the measures of fit proposed in this paper are based on the solution to this minimization problem, it is useful to discuss the problem and its solution in detail. Rather than move directly to the solution of the general problem, we shall first solve two simpler problems. This helps develop intuition for the general solution. In the first problem, xt and yt are serially uncorrelated scalars, and the representation follows by inspection. In the second problem, xt and yt are serially uncorrelated n×1 vectors, and the solution is slightly more difficult to derive. Finally, in the last problem, xt and yt are allowed to be serially correlated. Model 1 Suppose that xt, yt, and ut are scalar serially uncorrelated random variables. The problem is to choose σ xy to minimize the variance of u t , subject to the constraint that the covariance matrix of xt and yt remains positive semidefinite, that is By inspection, the solution sets σxy=σxσy and yields as the minimum. Since σxy=σxσy, xt and yt are perfectly correlated with (3) where γ=σx/σy. Equation (3) is important because it shows how to calculate fitted values of xt, given data on yt. Variants of equation (3) will hold for all the models considered. In each model, the minimum approximation error representation makes {xt} perfectly correlated with {y t}. In each model, the analogue of (3) provides a formula for calculating the fitted values of the variables in the model, given data from the actual economy. Model 2 Now suppose that xt and yt are serially uncorrelated random vectors with covariance matrices Σx and Σy, respectively. Let Σu=Σx+Σy -Σxy-Σyx denote the covariance matrix of ut. Since Σu is a matrix, there is not a unique way © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors MEASURES OF FIT 1018 309 JOURNAL OF POLITICAL ECONOMY to judge how “small” it is. A convenient measure of the size of ut is the where Σu,ij denotes the ijth element of Σu. trace of Σu, While convenient, this measure is not always ideal since it weights all variables equally. Below, we shall find a representation that minimizes tr(WΣu), where W is a prespecified n×n matrix. When all variables are equally important, W=In, and unequal weighting can be implemented by making W diagonal with the desired weights as the diagonal elements. The matrix W can also be used to focus attention on specific linear combinations of the variables that may be particularly interesting. For example, let G denote an n×n matrix and suppose that the researcher is primarily interested in the variables Gx t and Gy t . Then since tr(GΣuG’)=tr(G’GΣu), W can be chosen as G’G. The problem then is to choose Σxy to minimize tr(WΣu) subject to the is positive semidefinite. constraint that the covariance matrix of The solution is given below for the case in which Σx has rank k≤n. This occurs, for example, in economic models in which the number of variables exceeds the number of shocks. The solution is summarized in the following proposition. PROPOSITION. Assume (i) rank(Σx)=k≤n, (ii) rank(WΣxW’)= rank(Σx), and (iii) rank(Σy)=n. Let Cy denote an arbitrary n×n matrix square and let Cx denote an arbitrary n×k matrix root of Σy Let USV’ denote the singular value square root of Σx where U is an n×k orthogonal matrix (U’U=Ik), decomposition of S is a k×k diagonal matrix, and V is a k×k orthonormal matrix. Then is the unique matrix that minimizes tr(WΣu) subject to is positive semidefinite. the constraint that the covariance matrix of The proof is given in the Appendix. One important implication of this solution is that, like the scalar example, is singular and xt can be represented as the joint covariance matrix (4) where (Since U’U=W=Ik, this simplifies to the scalar result when xt and yt are scalars.) Model 3 This same approach can be used in a dynamic multivariate model with slight modifications; when ut is serially correlated, the weighted trace of the spectral density matrix rather than the covariance matrix can be minimized. To motivate the approach, it is useful to use the Cramer representations for xt, yt, and ut (see, e.g., Brillinger 1981, sec. 4.6). Assume that xt, yt, © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 310 CELEBRATION MEASURES OF FIT 1019 and ut are jointly covariance stationary with mean zero; the Cramer representation can be written as (5) where dz(ω)=(dz x(ω)′ dzy(ω)′ dzu(ω)′)′ is a complex valued vector of orthogonal increments, with where δ(ω-γ) is the dirac delta and S(ω) is the spectral density matrix of at frequency ω. Equation (5) represents xt, yt, and ut as the integral (sum) of increments dzx(ω), dzy(ω), and dzu(ω), which are uncorrelated across frequencies and have variances and covariances given by the spectra and cross spectra of xt, yt, and ut. Since the spectra are proportional to the is proportional to Ax(e-iω), ACGFs evaluated at z= e-iω, -iω is proportional to Axy(e ), and so forth. Now consider the problem of choosing Axy(z) to minimize the variance of ut. Since ut can be written as the integral of the uncorrelated increments dzu(ω), the variance of ut can be minimized by minimizing the variance of dzu(ω) for each ω. Since the increments are uncorrelated across frequency, the minimization problems can be solved independently for each frequency. Thus the analysis carried out for model 2 carries over directly, with spectral density matrices replacing covariance matrices. The minimum trace problem for model 2 is now solved frequency by frequency using the spectral density matrix. Like models 1–2, the solution yields (6) where Γ(ω) is the complex analogue of Γ from (4). Equation (6) implies (7) and (8) The autocovariances of ut follow directly from (8). Moreover, since dzx(ω) and dzy(ω) are perfectly correlated from (7), xt can be represented as a function of leads and lags of yt: (9) © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors MEASURES OF FIT 1020 311 JOURNAL OF POLITICAL ECONOMY where with xt can be calculated from leads and lags of yt. Thus fitted values of An Example The model considered in the next section describes the dynamic properties of output, consumption, investment, and employment as functions of a single productivity shock. To demonstrate the mechanics of the minimum approximation error representation for that model, assume that xt and yt are n×1 vectors and that xt is driven by a single iid(0, 1) shock ∊t: (10) where α(L) is an n×1 matrix polynomial in the lag operator, L. Assume that the Wold representation for the data is given by (11) where ⌰(L) is an n×n matrix polynomial in L, and et is an n×1 serially uncorrelated vector with mean zero and identity covariance matrix. The minimum error representation can then be computed directly from the matrix expressions given in the proposition. From (10), Ax(z) =α(z)α(zl )’ and, from (11), Αy(z)=⌰(z)⌰(z-1)’. Suppose that the weighting matrix is W=In, so that the trace of the spectral density of ut is to be minimized for α(e-iω) each frequency. In terms of the matrices in the proposition, Cx(ω)=α -iω and Cy(ω)=⌰(e ). Thus the cross spectrum/cross ACGF for xt and yt is chosen as Axy(e-iω)= α(e-iω)V(ω)U(ω)’⌰(eiω)’, where U(ω)S(ω)V(ω)’ is the ␣ (e-iω). (Since U(ω) and V(ω) are singular value decomposition of ⌰(eiω)’␣ complex matrices, V(ω)’ and U(ω)’ denote the transpose conjugates of V(ω) and U(ω), respectively.) The ACGF for ut follows from Au(e-iω)= Ax(e-iω)+Ay(e-iω)-Axy(e-iω)-Ayx(e-iω). Finally, to compute fitted values of xt from α (e the y t realization, note that dz x(ω)=Γ(ω)dz y(ω), where Γ(ω)=α iω -iω -1 )V(ω)U(ω)’⌰(e ) . Relative Mean Square Approximation Error A bound on the relative mean square approximation error for the economic model can be calculated directly from (8). The bound—analogous to a lower bound on 1-R2 from a regression—is (12) © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 312 CELEBRATION MEASURES OF FIT 1021 where [Au(z)]jj; and [Ay(z)]jj are the jth diagonal elements of Au(z) and Ay(z), respectively. Thus rj(ω) is the variance of the jth component of dzu(ω) relative to the jth component of dzy(ω), that is, the variance of the error relative to the variance of the data for each frequency. A plot of rj(ω) against frequency shows how well the economic model fits the data over different frequencies. Integrating the numerator and denominator of rj(ω) provides an overall measure of fit. (Note that since ut and xt are correlated, rj(ω) can be larger than one; i.e., the R2 of the model can be negative.)3 One advantage of rj(ω) is that it is unaffected by time-invariant linear filters applied to the variables. Filtering merely multiplies both the numerator and denominator of rj(ω) by the same constant, the squared gain of the filter. So, for example, rj(ω) is invariant to “Hod-rick-Prescott” filtering (see Hodrick and Prescott 1980; King and Rebelo 1993) or standard linear seasonal adjustment filters.4 The integrated version of the relative mean square approximation error is not invariant to filtering since it is a ratio of averages of both the numerator and denominator across frequencies. When the data are filtered, the integrated version of rj(ω) changes because the weights implicit in the averaging change. Frequencies for which the filter has a large gain are weighted more heavily than frequencies with a small gain. III. Measures of Fit for a Real Business Cycle Model In this section, a standard real business cycle model is evaluated using the measures of fit developed in the last section. The model, which derives from Kydland and Prescott (1982), is the “baseline” model of King, Plosser, and Rebelo (1988b). It is a one-sector neoclassical growth model driven by an exogenous stochastic trend in technology.5 3 The measure rj(ω) is not technically a metric since it does not satisfy the triangle inequality. 4 Standard seasonal adjustment filters such as the linear approximation to Census X-11 have zeros at the seasonal frequencies, so that rj(ω) is undefined at these frequencies for the filtered data. 5 This model is broadly similar to the model analyzed in Kydland and Prescott (1982). While the baseline model does not include the complications of time to build, inventories, time-nonseparable utility, and a transitory component to technology contained in the original Kydland and Prescott model, these complications have been shown to be reasonably unimportant for the empirical predictions of the model (see Hansen 1985). Moreover, the King, Plosser, and Rebelo baseline model appears to fit the data better at the very low frequencies than the original Kydland and Prescott model since it incorporates a stochastic trend rather than the deterministic trend present in the Kydland and Prescott formulation. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors MEASURES OF FIT 1022 313 JOURNAL OF POLITICAL ECONOMY This baseline model is analyzed, rather than a more complicated variant, for several reasons. First, the calibration/simulation exercises reported in King, Plosser, and Rebelo suggest that the model explains the relative variability of aggregate output, consumption, and investment, and it produces series with serial correlation properties broadly similar to the serial correlation properties of postwar U.S. data. Second, King, Plosser, Stock, and Watson (1991) show that the low-frequency/cointegration implications of the model are broadly consistent with similar postwar U.S. data. Finally, an understanding of where this baseline model fits the data and where it does not fit may suggest how the model should be modified. Only a brief sketch of the model is presented; a thorough discussion is contained in King, Plosser, and Rebelo (1988a, 1988b). The details of the model are as follows: where Ct denotes consumption, Lt is leisure, Qt is output, Kt is capital, Nt is employment, It is investment, and At is the stock of technology, with log(At) assumed to follow a random walk with drift γa and innovation εt. To analyze the model’s empirical predictions, the equilibrium of the model must be calculated as a function of the parameters β, θ, α, γa, and δ. This equilibrium implies a stochastic process for the variables Ct, Lt, Nt, Kt, It, and Qt, and these stochastic processes can then be compared to the stochastic processes characterizing U.S. postwar data. As is well known, the equilibrium can be calculated by maximizing the representative agent’s utility function subject to the technology and the resource constraints. In general, a closed-form expression for the equilibrium does not exist, and numerical methods must be used to calculate the stochastic process for the variables corresponding to the equilibrium. A variety of numerical approximations have been proposed (see Taylor and Uhlig [1990] for a survey); here I use the log linearization of the Euler equations proposed by King, Plosser, and Rebelo (1987). A formal justification for approximating © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 314 CELEBRATION MEASURES OF FIT 1023 the equilibrium of this stochastic nonlinear model near its deterministic steady state using linear methods is provided in Woodford (1986, theorem 2). The approximate solution yields a vector autoregression (VAR) for the logarithms of Qt,Ct,Kt, It, and Nt. (As in the standard convention, these logarithms will be denoted by lowercase letters.) All of the variables except nt are nonstationary but can be represented as stationary deviations about at, the logarithm of the stock of technology, which by assumption follows an integrated process. Thus qt, ct, it, and kt are cointegrated with a single common trend, a t. Indeed, not only are the variables in the VAR cointegrated, they are singular; the singularity follows since εt is the only shock to the system. The coefficients in the VAR are complicated functions and δ. Values for these of the structural parameters β, θ, α, γa, parameters are the same as those used by King, Plosser, and Rebelo (1988b): when the variables are measured quarterly, the parameter values are α=.58, δ=.025, γa= .004, σε=.010, and β=.988, and θ is chosen so that the steadystate value of N is .20. These parameter values were chosen so that the model’s steady-state behavior matches postwar U.S. data.6 With these values for the parameters, the VAR describing the equilibrium can be calculated and the ACGF of xt=(Δqt Δct Δit nt)’ follows directly.7 These autocovariances will be compared to the autocovariances of postwar data for the United States. The data used here are the same data used by King, Plosser, Stock, and Watson (1991). The output measure is total real private GNP, defined as total real GNP less government purchases of goods and services. The measure of consumption is total real consumption expenditures, and the measure of investment is total real fixed investment. The measure of employment is total labor hours in private nonagricultural establishments. All variables are expressed in per capita terms using the total civilian noninstitutional 6 The choice of parameter values is described in King, Plosser, and Rebelo (1988a). The value of a was chosen to equal the average value of labor’s share of gross national product over 1948–86. The value of γa was chosen as the common average quarterly rate of growth of per capita values of real GNP, consumption of nondurables and services, and gross fixed investment. The depreciation rate was chosen to yield a gross investment share of GNP of approximately 30 percent. The parameter θ was chosen so that the model’s steady-state value of N matched the average workweek as a fraction of total hours over 1948–86. The discaount rate β was chosen so that the model’s steady-state annual interest rate matched the average rate of return on equity over 1948–81. The value of σε=.01 appears to have been chosen as a convenient normalization. This value is used here because it does a remarkably good job matching the very low frequency movements in output, consumption, and investment. 7 Of course, this not the only possible definition of xt. The only restriction on xt is covariance stationarity, so, e.g., ct-qt and it-qt could be included as elements. 8 All data are taken from Citibase. With the Citibase labels, the precise variables used were gnp82—gge82 for output, gc82 for consumption, and gif82 for investment. The © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors MEASURES OF FIT 1024 315 JOURNAL OF POLITICAL ECONOMY population over the age of 16.8 Let denote the logarithm of per capita private output, the logarithm of per capita consumption expenditures, and so forth. Then the data used in the analysis can be written as The analysis presented in the last section assumed that the ACGF/ spectrum of yt was known. In practice, of course, this is not the case, and the spectrum must be estimated. In this work, the spectrum of yt was estimated in two different ways. First, an autoregressive spectral estimator was used, calculated by first estimating a VAR for the variables and then forming the implied spectral density matrix. As in King, Plosser, Stock, and Watson (1991), the VAR was estimated imposing a cointegration constraint between output, consumption, and investment. Thus the VAR onto a was specified as the regression of constant and six lags of wt. The parameters of the VAR were estimated using data for 1950–88. (Values before 1950 were used as lags in the regression for the initial observations.) Second, a standard nonparametric spectral estimator was also calculated. The spectrum was estimated by a simple average of 10 periodogram ordinates after prewhitening employment with the filter 1—.95L. These two estimators yielded similar values for the measures of fit, and to conserve space only the results for the autoregressive spectral estimator are reported. For each variable, figure 1 presents the spectrum implied by the model, the spectrum of the data, and the spectrum of the error required to reconcile the model with the data.9 The error process was chosen to minimize the unweighted trace of the error spectral density matrix, subject to the positive semidefiniteness constraint discussed in the last section. Thus the objective function weighted all the variables equally. For output, consumption, and investment, the model and data spectra differ little for very low frequencies (periods greater than 50 quarters) and for output and investment at high frequencies (periods less than five quarters). There are significant differences between the model and data spectra for periods typically associated with the business cycle; the largest differences occur at a frequency corresponding to approximately 20 quarters. The spectra of Δnt and are quite different. In addition to large differences at business cycle frequencies, the spectra are also very different at low frequencies. measure of total labor hours was constructed as total employment in nonagricultural establishments (lhem) less total government employment (lpgov) multiplied by average weekly hours (lhch). The population series was P16. 9 Figure 1 is reminiscent of figures in Howrey (1971, 1972), who calculated the spectra implied by the Klein-Goldberger and Wharton models. A similar exercise is carried out in Soderlind (1993), who compares the spectra of variables in the KydlandPrescott model to the spectra of postwar U.S. data. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 316 CELEBRATION MEASURES OF FIT 1025 The model implies that employment is stationary so that its growth rate has a spectrum that vanishes at frequency zero. In contrast, the data suggest significant low-frequency variation in postwar U.S. employment.10 The figure shows that relatively little error is needed to reconcile the model and the data for output, consumption, and investment over the very low frequencies. On the other hand, error with a variance on the order of 40–50 percent of the magnitude of the variance of the series is necessary for the components of output, consumption, and investment with periods in the 6–32-quarter range. At higher frequencies, the model is able to match the stochastic process describing investment, but not the processes describing the other series. Table 1 provides a summary of the relative mean square approximation error (RMSAE) for a variety of weighting functions and filters. Each panel shows the RMSAE for the variables for five different minimum error representations. Column 1 presents results for the representation that obtains when the unweighted trace of the spectrum is minimized; this is the representation used to construct the error spectra shown in figure 1. Column 2 summarizes the results for the representation that minimizes the output error, with no weight placed on the other variables. Column 3 summarizes results for the representation that minimizes the consumption error, and so forth. Panel A presents the results for the differences of the data integrated across all frequencies, panel B shows results for the levels of the series detrended by the Hodrick-Prescott filter integrated across all frequencies, and panel C presents results for the levels of the series integrated over business cycle frequencies (6–32 quarters). The trade-off inherent in the different representations is evident in all panels. For example, in panel A, with the minimum output error representation, the RMSAE for output is 26 percent, and the RMSAE for consumption is 78 percent; when the minimum consumption error representation is chosen, the RMSAE for consumption falls to 21 percent but the RMSAE for output rises to 75 percent. When all the variables are equally weighted, the RMSAE is 52 percent for output and 66 percent for consumption. Panel C shows that most of this trade-off occurs at the high frequencies, at least for output, consumption, and investment; over the business cycle frequencies their RMSAEs are in the 40–60 percent range. As explained in Section II, fitted values of the model’s variables can be constructed using the minimum error representation together with the 10 The figures do not include standard errors for the spectra estimated from these data. These standard errors are large—approximately one-third the size of the estimated spectra. The standard errors for the RMSAE, averaged across frequencies, are considerably smaller. These are included in tables 1 and 2 below. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors MEASURES OF FIT 317 FIG. 1.—Decomposition of spectra: a, output; b, consumption; c, investment; d, employment. Dotted lines refer to the data, dashed lines to the model, and solid lines to approximation error. 1026 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 318 CELEBRATION FIG. 1.—Continued 1027 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors TAB LE 1 RELATIVE MEAN SQUARE APPROXIMATION ERROR NOTE.—Relative mean square approximation error is the lower bound of the variance of the approximation error divided by the variance of the series. Each column represents the relative mean square approximation error of the row variable constructed from the representation that minimizes the weighted trace of the error spectrum. The weights are summarized by the column headings. For example, col. 1 is the equally weighted trace, col. 2 puts all the weight on the output error, etc. The numbers in parentheses are standard errors based on the sampling error in the estimated VAR coefficients used to estimate the data spectrum. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 320 CELEBRATION MEASURES OF FIT 1029 observed data. Since the measurement error model represents yt as xt plus error, the standard signal extraction formula can be used to extract {xt} from {yt}. In general, of course, signal extraction methods will yield an that is not exact in the sense that estimate of xt, say In the present context, the estimate will be exact since the measurement error process is chosen so that dzx(ω) and dzy(ω) are perfectly correlated for all ω.11 Figure 2 shows the realizations of the data and the realizations of the variables in the model calculated from the data using the equally weighted minimum output error representation.12 In figure 2a, which shows the results for output, the model seems capable of capturing the long swings in the postwar U.S. data but not capable of capturing the cyclical variability in the data. Private per capita GNP fell by 8.4 percent from the cyclical peak in 1973 to the trough in 1975 and by 6.8 percent from the peak in 1979 to the trough in 1982. In contrast, the corresponding drops in Qt—output in the model—were 3.1 percent and 3.0 percent, respectively. The dampened cyclical swings in consumption and investment, shown in figure 2b and c, are even more dramatic. Finally, figure 2d shows that the model predicts changes in employment that have little to do with the changes observed in the United States during the postwar period.13 One possible explanation for the relatively poor fit of the model is that the “calibrated” values of the parameters are wrong. In particular, Christiano and Eichenbaum (1990) show that the model’s predictions change in an important way when the technology process changes from a random walk to a stationary AR(1). Table 2 shows how the model fares for a range of values of the AR(1) coefficient for technology, denoted by ρa. Panel A of the table shows the results for first differences of the variables across all frequencies, panel B presents results for the Hodrick-Prescott detrended levels of the series, and panel C shows the results for the levels 11 More precisely, the estimate is exact in the sense that converges in mean square to xt as j→∞. 12 As shown in eq. (9), xt can be calculated as β(L)yt, where β(L) is the inverse Fourier transform of G(ω). To calculate the estimates shown in the figure, Γ(ω) was calculated at 128 equally spaced frequencies between zero and π. Since β(L) is twosided, pre-and postsample values of yt are required to form β(L)xt. These pre- and postsample values were replaced with the sample means of the yt data. The first differences, xt and yt, were then accumulated to form the levels series shown in the figure. 13 The calculations required to construct figs. 1 and 2 and the results in table 1 are easily carried out. For this example, the model spectrum, data spectrum, RMSAEs, and fitted values were calculated in less than a minute on a standard desktop computer. A GAUSS program for these calculations is available from the author. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors MEASURES OF FIT 321 FIG. 2.—Data: a, output; b, consumption; c, investment; d, employment. Solid lines refer to U.S. data and dashed lines to realizations from the model. 1030 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 322 CELEBRATION FIG. 2—Continued 1031 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors MEASURES OF FIT 1032 323 JOURNAL OF POLITICAL ECONOMY TAB LE 2 RELATIVE MEAN SQUARE APPROXIMATION ERROR AS A FUNCTION OF THE AR(1) COEFFICIENT FOR TECHNOLOGY NOTE.—Relative mean square approximation error is the lower bound of the variance of the approximation error divided by the variance of the series. Each column represents the relative mean square approximation error of the row variable constructed from the representation that minimizes the weighted trace of the error spectrum. The column headings represent the AR(1) coefficient for the process of the logarithm of productivity in the model. For example, col. 1 represents results for the model with random walk technological progress. The numbers in parentheses are standard errors based on the sampling error in the estimated VAR coefficients used to estimate the data spectrum. of the series over the “business cycle” frequencies. From panel A, the value of ρa has little effect on the fit of the model averaged across all frequencies. In particular, as ρα falls from 1.0 to .90, the RMSAE increases slightly for consumption, falls for investment, and changes little for output and employment. In contrast, the value of ρa has a significant effect on the fit of the model over business cycle frequencies. For example, panel C shows that as ρa falls from 1.0 to .90, the RMSAE falls for output (.43 to .33), for investment (.42 to .17), and for employment (.72 to .52); it increases for consumption (.52 to .66). The source of the changes in the RMSAEs can be seen in figure 3, which plots the spectra of the variables in models with ρa=1 and ρa=.90. The largest difference between the spectra of the models is the increase in variance in output, investment, and employment as ρa falls from 1.0 to ρa.90. The economic mechanism behind this increased variance is the increase in intertemporal substitution in response to a technology shock. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 324 CELEBRATION MEASURES OF FIT 1033 When ρa=.90, technology shocks are transitory, and they can be exploited only by large transitory increases in employment and investment. It is interesting to note that while this mechanism increases the variance of the growth rates in employment, investment, and output, it has little effect on their autocorrelations. That is, as ρα changes from 1.0 to .90, the shape of the spectra changes little. Before we leave this section, six additional points deserve mention. First, the fitted values in figure 2 are quantitatively and conceptually similar to figures presented in Christiano (1988) and Plosser (1989). They calculated the Solow residual from actual data and then simulated the economic model using this residual as the forcing process. Implicitly, they assumed that the model and data were the same in the terms of their Solow residual, and then asked whether the model and data were similar in other dimensions. Figure 2 is constructed by making the model and data as close as possible in one dimension (in this case the trace of the variance of the implied approximation error) and then asking whether the model and data are similar in other dimensions. The difference between the two approaches can be highlighted by considering the circumstances in which they would produce exactly the same figure. If the Solow residual computed from the actual data followed exactly the same stochastic process as the change in productivity in the model, and if the approximation error representation was constructed by minimizing the variance of the difference between the Solow residual in the data and productivity growth in the model, then the two figures would be identical. Thus the figures will differ if the stochastic process for the empirical Solow residual is not the same as assumed in the model, or the approximation error representation is chosen to make the model and data close in some dimension other than productivity growth. Second, the inability of the model to capture the business cycle properties of the data is not an artifact of the minimum measurement error representation used to form the projection of xt onto yτ, τ = 1, …, n. Rather, it follows directly from a comparison of the spectra of xt and yt. The fitted values are constrained to have an ACGF/ spectra given by the economic model. Figure 1 shows that, for all the variables, the spectral power over the business cycle frequencies is significantly less for the model than for the data. Therefore, fitted values from the model are constrained to have less cyclical variability than the data. Third, the ability of the model to mimic the behavior of the data depends critically on the size of the variance of the technology shock. The value of σε used in the analysis is two and one-half times larger than the drift in the series. Thus if the εt were approximately normally distributed, the stock of technology At would fall in one out of three quarters on average. Reducing © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors MEASURES OF FIT 325 FIG. 3.—Data and model spectra: a, output; b, consumption; c, investment; d, employment. Dotted lines refer to the data, solid lines to the model with ρα=1.00, and dashed lines to the model with ρα=.90. 1034 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 326 CELEBRATION FIG.3—Continued 1035 © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors MEASURES OF FIT 1036 327 JOURNAL OF POLITICAL ECONOMY the standard deviation of the technology shock so that it equals the average growth in at drastically increases the size of the measurement error necessary to reconcile the model with the data. For example, integrated across all frequencies, the RMSAE for output increases from 52 percent to 74 percent. Fourth, there is nothing inherent in the structure of the model that precludes the use of classical statistical procedures. Altug (1989) used maximum likelihood methods to study a version of the model that is augmented with serially correlated classical measurement errors. Singleton (1988) and Christiano and Eichenbaum (1992) pointed out that generalized method of moments procedures can be used to analyze moment implications of models like the one presented above. In the empirical work of Christiano and Eichenbaum the singularity in the probability density function of the data that is implied by the model was finessed in two ways. First, limited information estimation and testing methods were used, and second, the authors assumed that their data on employment were measured with error. Fifth, many if not all of the empirical shortcomings of this model have been noted by other researchers. King, Plosser, and Rebelo clearly show that the model is not capable of explaining the variation in employment that is observed in the actual data. The implausibility of the large technology shocks is discussed in detail in Summers (1986), Mankiw (1989), and McCallum (1989). Finally, the analysis above has concentrated on the ability of the model to explain the variability in output, consumption, investment, and employment across different frequencies. While it is possible to analyze the covariation of these series using the cross spectrum of the measurement error, such an analysis has not been carried out here. This is a particularly important omission since this is the dimension in which the baseline real business cycle model is typically thought to fail. For example, Christiano and Eichenbaum (1992) and Rotemberg and Woodford (1992) use the model’s counterfactual implication of a high correlation between average productivity and output growth as starting points for their analysis, and the empirical literature on the intertemporal capital asset pricing model beginning with Hansen and Singleton (1982) suggests that the asset pricing implications of the model are inconsistent with the data. It would be useful to derive simple summary statistics based on the cross spectra of the measurement error and the data to highlight the ability of the model to explain covariation among the series. IV. Discussion The discussion thus far has assumed that the parameter values of the economic model are known. A natural question is whether the measures © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 328 CELEBRATION MEASURES OF FIT 1037 of fit discussed in this paper can form the basis for estimators of these parameters. Does it make sense, for example, to estimate unknown parameters by minimizing some function of the relative mean square error, rj(ω) given in equation (12)? This certainly seems sensible. For example, a researcher may want to “calibrate” his model with a value ofρα=.90 rather than 1.0, because this value produces spectra closer to the estimated spectra of data over the business cycle frequencies. Yet dropping the standard statistical assumption that the economic model is correctly specified raises a number of important issues. Foremost among these is the meaning of the parameters. If the model does not necessarily describe the data, then what do the parameters measure? Presumably, the model is meant to describe certain characteristics of the data’s stochastic process (the business cycle or the growth properties, for example), while ignoring other characteristics. It then makes sense to define the model’s parameters as those that minimize the differences between the model and the data’s stochastic process in dimensions that the model is attempting to explain. So, for example, it seems sensible to define the parameters of a growth model as those that minimize rj(ω) over very low frequencies, or to define the parameters of a business cycle model as those that minimize rj(ω) over business cycle frequencies. Given this definition of the parameters, constructing an analog estimator (see Manski 1988) by minimizing corresponds to a standard statistical practice. Of course, the parameters may also be defined using other characteristics of the model and the stochastic process describing the data. For example, in standard “calibration” estimation exercises, many of the parameters are implicitly defined in terms of first moments of the data. Parameters are chosen so that the first moments of the variables in the model’s steady state match the first moments of the data. Two final points deserve mention. First, since the measures of fit developed in this paper are based on a representation that minimizes the discrepancy between the model and the data, they serve only as a bound on the fit of the model. Models with large RMSAEs do not fit the data well. Models with small RMSAEs fit the data well given certain assumptions about the correlation properties of the noise that separates the model and the data, but may fit the data poorly given other assumptions about the noise. Finally, while this paper has concentrated on measures of fit motivated by a model of measurement error, other measures are certainly possible. For example, one measure, which like the measures in this paper uses only the autocovariances implied by the model and the data, is the expected log likelihood ratio using the normal probability density function (pdf) of the data and the model. More precisely, if g(x) denotes the normal pdf © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors MEASURES OF FIT 1038 329 JOURNAL OF POLITICAL ECONOMY constructed from the autocovariances of the data, f(x) denotes the normal pdf constructed for the autocovariances implied by the model, and Eg is the expectation operator taken with respect to g(x), the expected log likelihood ratio I(g, f)= Eg{log[g(x)/f(x)]} can be used to measure the distance between the densities f(·) and g(·); I(g, f) is the Kullback-Leibler information criterion (KLIC), which plays an important role in the statistical literature on model selection (e.g., Akaike 1973) and quasi-maximum likelihood estimation (White 1982). Unfortunately, the KLIC will not be defined when f(x) is singular and g(x) is not; the KLIC distance between the two densities is infinite. Thus, for example, it would add no additional information on the fit of the real business cycle model analyzed in Section III beyond pointing out the singularity. Arguably, one of the most informative diagnostics presented in this paper is the plot of the model and data spectra. For example, figures 1 and 2 show that the data spectra have mass concentrated around the business cycle frequencies, but the model spectra do not. Any metric comparing the data and model spectra may serve as a useful measure of fit. The RMSAE proposed here has the advantage that it can be interpreted like 1R2 from a regression, but any summary statistic discards potentially useful information contained in plots such as figures 1 and 2. Some practical advice, therefore, is to present both model and data spectra as a convenient way of comparing their complete set of second moments. Appendix To prove the proposition, first parameterize Σx, Σy, and Σxy as (A1) (A2) (A3) where Cx is n×k with full column rank, G is n×k, and Σ is positive semidefinite. Since ΣU=Σx+Σy-Σxy-Σyx, minimizing tr(WΣu) with Σx and Σy given is equivalent to maximizing tr(WΣxy)=tr(WCxG’). Given an arbitrary factorization of Σx of the form (A1), the problem is to find the n×k matrix G to maximize tr(WCxG’) subject to the constraint that Σy- GG´=Σ is positive semidefinite. Then Σy-GG´ is positive semidefinite if and only if In´ is positive Let semidefinite, which in turn is true if and only if all the eigenvalues of GG’ are less than or ´ are the same as those of , the problem can equal to one. Since the eigenvalues of be written as (A4) © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 330 CELEBRATION MEASURES OF FIT where λ i( ) denotes the ith eigenvalue of tr(AB)=tr(BA) for conformable matrices A and B. 1039 , and I have used the fact that Let QDR´ denote the singular value decomposition of , where Q is an n ×k orthogonal matrix, R is a k×k orthonormal matrix, and D is a k×k diagonal matrix with elements dij. Since and since the solution to (A4) is seen to require that λi( )=1, i=1, …, k. This implies that =I. Write the singular value decomposition of as USV´; then where Since the maximization problem can be written as (A5) Assumptions i-iii of the proposition imply that has full column rank so that S is a diagonal matrix with strictly positive diagonal elements. Thus since U´U=Ik, the maximization is achieved by Working backward, we see that G=C y UV’, so that Uniqueness follows since this choice of Σxy does not depend on the (arbitrary) choice of the matrix square roots, Cx and Cy. To see this, let y and x denote other matrix square roots of Σy and Σx. Then y=CyRy and x =CxRx, where Ry and Rx are orthonormal matrices. From the analysis above, this yields where is the singular value decomposition of . By inspection, and so that References Akaike, H. “Information Theory and an Extension of the Maximum Likelihood Principle.” In Proceedings of the Second International Symposium on Information Theory, edited by B.N.Petrov and F.Csaki. Budapest: Akademiai Kaido, 1973. Altug, Sumru. “Time-to-Build and Aggregate Fluctuations: Some New Evidence.” Internal. Econ. Rev. 30 (November 1989): 889–920. Brillinger, David R. Time Series: Data Analysis and Theory. San Francisco: Holden-Day, 1981. Campbell, John Y., and Shiller, Robert J. “The Dividend-Price Ratio and Expectations of Future Dividends and Discount Factors.” Rev. Financial Studies 1, no. 3 (1988): 195–228. Christiano, Lawrence J. “Why Does Inventory Investment Fluctuate So Much?” J.Monetary Econ. 21 (March/May 1988): 247–80. Christiano, Lawrence J., and Eichenbaum, Martin. “Unit Roots in Real GNP: Do We Know and Do We Care?” Carnegie-Rochester Conf. Ser. Public Policy 32 (Spring 1990): 7–61. ————. “Current Real-Business-Cycle Theories and Aggregate Labor-Market Fluctuations.” A.E.R. 82 (June 1992): 430–50. Cochrane, John H. “Explaining the Variance of Price-Dividend Ratios.” Rev. Financial Studies 5, no. 2 (1992): 243–80. Durlauf, Steven N., and Hall, Robert E. “Measuring Noise in Stock Prices.” Manuscript. Stanford, Calif.: Stanford Univ., 1989. Geweke, John. “The Dynamic Factor Analysis of Economic Time-Series Models.” In Latent Variables in Socio-economic Models, edited by Dennis J.Aigner and Arthur S.Goldberger. Amsterdam: North-Holland, 1977. Haavelmo, Trygve. “The Probability Approach in Econometrics.” Econometrica 12 (suppl.; July 1944): 1–115. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors MEASURES OF FIT 1040 331 JOURNAL OF POLITICAL ECONOMY Hansen, Gary D. “Indivisible Labor and the Business Cycle.” J.Monetary Econ. 16 (November 1985): 309–27. Hansen, Gary D., and Sargent, Thomas J. “Straight Time and Overtime in Equilibrium.” J.Monetary Econ. 21 (March/May 1988): 281–308. Hansen, Lars Peter, and Jagannathan, Ravi. “Implications of Security Market Data for Models of Dynamic Economies.” J.P.E. 99 (April 1991): 255–62. Hansen, Lars Peter, and Sargent, Thomas J. “Formulating and Estimating Dynamic Linear Rational Expectations Models.” J.Econ. Dynamics and Control 2 (February 1980): 7–46. Hansen, Lars Peter, and Singleton, Kenneth J. “Generalized Instrumental Variables Estimation of Nonlinear Rational Expectations Models.” Econometrica 50 (September 1982): 1269–86. Hodrick, Robert J., and Prescott, Edward C. “Post-War U.S. Business Cycles: An Empirical Investigation.” Manuscript. Pittsburgh: Carnegie Mellon Univ., 1980. Howrey, E.Philip. “Stochastic Properties of the Klein-Goldberger Model.” Econometrica 39 (January 1971): 73–87. ——. “Dynamic Properties of a Condensed Version of the Wharton Model.” In Econometric Models of Cyclical Behavior, vol. 2, edited by Bert G. Hickman. New York: Columbia Univ. Press (for NBER), 1972. King, Robert G.; Plosser, Charles I.; and Rebelo, Sergio T. “Production, Growth, and Business Cycles: Technical Appendix.” Manuscript. Rochester, N.Y.: Univ. Rochester, 1987. ——. “Production, Growth and Business Cycles: I. The Basic Neoclassical Model.” J.Monetary Econ. 21 (March/May 1988): 195–232. (a) ——. “Production, Growth and Business Cycles: II. New Directions.” J. Monetary Econ. 21 (March/May 1988): 309–41. (b) King, Robert G.; Plosser, Charles I.; Stock, James H.; and Watson, Mark W. “Stochastic Trends and Economic Fluctuations.” A.E.R. 81 (September 1991): 819–40. King, Robert G., and Rebelo, Sergio T. “Low Frequency Filtering and Real Business Cycle Models.” J.Econ. Dynamics and Control (1993), in press. Kydland, Finn E., and Prescott, Edward C. “Time to Build and Aggregate Fluctuations.” Econometrica 50 (November 1982): 1345–70. McCallum, Bennett T. “Real Business Cycle Models.” In Modern Business Cycle Theory, edited by Robert J.Barro. Cambridge, Mass.: Harvard Univ. Press, 1989. McFadden, Daniel. “Econometric Models of Probabilistic Choice.” In Structural Analysis of Discrete Data with Econometric Applications, edited by Charles Manski and Daniel McFadden. Cambridge, Mass.: MIT Press, 1981. Mankiw, N.Gregory. “Real Business Cycles: A New Keynesian Perspective.” J.Econ. Perspectives 3 (Summer 1989): 79–90. Mankiw, N.Gregory; Runkle, David E.; and Shapiro, Matthew D. “Are Preliminary Announcements of the Money Stock Rational Forecasts?” J.Monetary Econ. 14 (July 1984): 15–27. Mankiw, N.Gregory, and Shapiro, Matthew D. “News or Noise: An Analysis of GNP Revisions.” Survey Current Bus. 66 (May 1986): 20–25. Manski, Charles F. Analog Estimation Methods in Econometrics. New York: Chapman and Hall, 1988. Plosser, Charles I. “Understanding Real Business Cycle Models.” J.Econ. Perspectives 3 (Summer 1989): 51–77. Prescott, Edward C. “Theory Ahead of Business-Cycle Measurement.” Carnegie-Rochester Conf. Ser. Public Policy 25 (Autumn 1986): 11–44. Rotemberg, Julio J., and Woodford, Michael. “Oligopolistic Pricing and the Effects of Aggregate Demand on Economic Activity.” J.P.E. 100 (December 1992): 1153–1207. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 332 CELEBRATION MEASURES OF FIT 1041 Sargent, Thomas J. “Two Models of Measurements and the Investment Accelerator.” J.P.E. 97 (April 1989): 251–87. Sargent, Thomas J., and Sims, Christopher A. “Business Cycle Modeling without Pretending to Have Too Much a Priori Economic Theory.” In New Methods in Business Cycle Research, by Christopher A.Sims et al. Minneapolis: Fed. Reserve Bank, 1977. Singleton, Kenneth J. “Econometric Issues in the Analysis of Equilibrium Business Cycle Models.” J.Monetary Econ. 21 (March/May 1988): 361–86. Soderlind, Paul. “Cyclical Properties of a Real Business Cycle Model.” Manuscript. Princeton, N.J.: Princeton Univ., Dept. Econ., 1993. Summers, Lawrence H. “Some Skeptical Observations on Real Business Cycle Theory.” Fed. Reserve Bank Minneapolis Q. Rev. 10 (Fall 1986): 23–27. Taylor, John B., and Uhlig, H. “Solving Nonlinear Stochastic Growth Models: A Comparison of Alternative Solution Methods.” J.Bus. and Econ. Statis. 8 (January 1990): 1–17. White, Halbert. “Maximum Likelihood Estimation of Misspecified Models.” Econometrica 50 (January 1982): 1–25. Woodford, Michael. “Stationary Sunspot Equilibria: The Case of Small Fluctuations around a Deterministic Steady State.” Manuscript. Chicago: Univ. Chicago, 1986. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors CHAPTER 18 333 JOURNAL OF APPLIED ECONOMETRICS, VOL. 9, S123–S144 (1994) STATISTICAL INFERENCE IN CALIBRATED MODELS FABIO CANOVA Department of Economics, Universitat Pompeu Fabra, Balmes 132, 08008 Barcelona, Spain and Department of Economics, Università di Catania, 95100 Catania, Italy, and CEPR SUMMARY This paper describes a Monte Carlo procedure to assess the performance of calibrated dynamic general equilibrium models. The procedure formalizes the choice of parameters and the evaluation of the model and provides an efficient way to conduct a sensitivity analysis for perturbations of the parameters within a reasonable range. As an illustration the methodology is applied to two problems: the equity premium puzzle and how much of the variance of actual US output is explained by a real business cycle model. 1. INTRODUCTION The current macroeconometrics literature has proposed two ways to confront general equilibrium rational expectations models with data. The first, an estimation approach, is the direct descendant of the econometric methodology proposed 50 years ago by Haavelmo (1944). The second, a calibration approach, finds its justification in the work of Frisch (1933) and is closely linked to the computable general equilibrium literature surveyed e.g. in Shoven and Whalley (1984). The two methodologies share the same strategy in terms of model specification and solution. Both approaches start from formulating a fully specified general equilibrium dynamic model and in selecting convenient functional forms for preferences, technology, and exogenous driving forces. They then proceed to find a decision rule for the endogenous variables in terms of the exogenous and predetermined variables (the states) and the parameters. When the model is nonlinear, closed-form expressions for the decision rules may not exist and both approaches rely on recent advantages in numerical methods to find an approximate solution which is valid either locally or globally (see e.g. the January 1990 issue of the Journal of Business and Economic Statistics for a survey of the methods and Christiano, 1990, and Dotsey and Mao, 1991, for a comparison of the accuracy of the approximations). It is when it comes to choosing the parameters to be used in the simulations and in evaluating the performance of the model that several differences emerge. The first procedure attempts to find the parameters of the decision rule that best fit the data either by maximum likelihood (ML) (see e.g. Hansen and Sargent, 1979, or Altug, 1989) or generalized method of moments (GMM) (see e.g. Hansen and Singleton, 1983, or Burnside et al., 1993). The validity of the specification is examined by testing restrictions, by general goodness of fit tests or by comparing the fit of two nested models. The second approach ‘calibrates’ parameters using a set of alternative rules which includes matching long-run averages, using previous microevidence or a priori selection, and assesses the fit of the model with an informal distance criterion. These differences are tightly linked to the questions the two approaches ask. Roughly speaking, the estimation approach asks the question ‘Given that the model is true, how false CCC 0883–7252/94/0S0S123–22 © 1994 by John Wiley & Sons, Ltd. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors Received July 1992 Revised August 1994 334 S124 CALIBRATION F. CANOVA is it?’ while the calibration approach asks ‘Given that the model is false, how true is it?’ Implicit in the process of estimation is in fact the belief that the probability structure of a model is sufficiently well specified to provide an accurate description of the data. Because economic models are built with an emphasis on tractability, they are often probabilistically underspecified so that measurement errors or unobservable shocks are added at the estimation stage to complete their probability structure (see e.g. Hansen and Sargent, 1980, or Altug, 1989). By testing the model, a researcher takes the model seriously as a datagenerating process (DGP) and examines what features of the specification are at variance with the data. A calibrationist takes the opposite view: the model, as a DGP for the data, is false. That is, as the sample size grows, it is known that the data are generated by the model will be at increasingly greater variance with the observed time series. An economic model is seen, at best, as an approximation to the true DGP which need not be either accurate or realistic and, as such, should not be regarded as a null hypothesis to be statistically tested (see Prescott, 1991, p. 5). In confronting the model with the data, a calibrationist wants to indicate the dimensions where the approximation is poor and suggest modifications to the theoretical model in order to obtain a better approximation. Both methodologies have weak points. Model estimation involves a degree of arbitrariness in specifying which variables are unobservable or measured with error. In the limit, since all variables are indeed measured with error, no estimation seems possible and fruitful. In addition, tests of the model’s restrictions may fail to indicate how to alter the specification to obtain a better fit. The limitations of the calibration approach are also well known. First, the selection criterion for parameters which do not measure long-run averages is informally specified and may lead to contradictory choices. Information used in different studies may in fact be inconsistent (e.g. a parameter chosen to match labour payments from firms in national account data may not equal the value chosen to match the labour income received by households) and the range of estimates for certain parameters (e.g. risk aversion parameter) is so large that selection biases may be important. Second, the outcomes of the simulations typically depend on the choice of unmeasured parameters. However, although some authors (see e.g. Prescott, 1991, p. 7, or Kydland, 1992, p. 478) regard a calibration exercise as incomplete unless the sensitivity of the results to reasonable perturbations of the parameters selected a priori or not well tied to the data is reported, such an analysis is not often performed. Third, because the degree of confidence in the results depends on both the degree of confidence in the theory and in the underlying measurement of the parameters and because either parameter uncertainty is disregarded or, when a search is undertaken, the number of replications typically performed is small, we must resort to informal techniques to judge the relevance of the theory. This paper attempts to eliminate some of the weaknesses of calibration procedures while maintaining the general analytical strategy employed in calibration exercises. The focus is on trying to formalize the selection of the parameters and the evaluation process and in designing procedures for meaningful robustness analysis on the outcomes of the simulations. The technique we propose shares features with those recently described by Gregory and Smith (1991) and Kwan (1990), has similarities with stochastic simulation techniques employed in dynamic nonlinear large scale macro models (see e.g. Fair, 1991), and generalizes techniques on randomized design for strata existing in the static computable general equilibrium literature (see e.g. Harrison and Vinod, 1989). The idea of the technique is simple. We would like to reproduce features of actual data, which is taken to be the realization of an unknown vector stochastic process, with an ‘artificial economy’ which is almost surely the incorrect generating mechanism for the actual data. The features we may be interested in include conditional and unconditional moments (or densities), the autocovariance function of the data, functions of these quantities (e.g. measures of relative volatility), or specific events (e.g. a recession). A model © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors STATISTICAL INFERENCE IN CALIBRATED MODELS STATISTICAL INFERENCE IN CALIBRATED MODELS 335 S125 is simulated repeatedly using a Monte Carlo procedure which randomizes over both the exogenous stochastic processes and the parameters. Parameters are drawn from a data-based density which is consistent with the information available to a simulator (which may include both time-series and cross-sectional aspects). We judge the validity of a model on its ability to reproduce a number of ‘stylized facts’ of the actual economy (see Friedman, 1959). The metric used to evaluate the discrepancy of the model from the data is probabilistic. We construct the simulated distribution of the statistics of interest and, taking the actual realization of the statistic as a critical value, examine (1) in what percentile of the simulated distribution the actual value lies and (2) how much of the simulated distribution is within a k% region centred around the critical value. Extreme values for the percentile (say, below α% or above (1-α)%) or a value smaller than k for the second probability indicates a poor fit in the dimensions examined. The approach has several appealing features. First, it accounts for the uncertainty faced by a simulator in choosing the parameters of the model in a realistic way. Second, it has a built-in feature for global sensitivity analysis on the support of the parameter space and allows for other forms of conditional or local sensitivity analysis. Third, it provides general evaluation criteria and a simple and convenient framework to judge the relevance of the theory. The paper is divided into six sections. The next section introduces the technique, provides a justification for the approach and describes the details involved in the implementation of the procedure. Section 3 deals with robustness analysis. Section 4 spells out the relationship with existing techniques. Two examples describing the potential of the technique for problems of practical interest appear in Section 5. Section 6 presents conclusions. 2. THE TECHNIQUE A General Framework of Analysis We assume that a researcher is faced with an m×1 vector of time series which are the realizations of a vector stochastic process and that she is interested in reproducing features of using a dynamic general equilibrium model. is assumed to have a continuous but unknown distribution and moments up to the nth. For the sake of presentation we assume is independent of t but shifts in the unconditional that the unconditional distribution of distribution of at known points can easily be handled. may include macro variables like GNP, consumption, interest rates, etc. We also assume that dynamic economic theory gives us a model expressing the endogenous variables Xt as a function of exogenous and predetermined variables Zt (the states of the problem) and of the parameters β. Zt may include objects like the existing capital stock, exogenous fiscal, and monetary variables or shocks to technologies and preferences. We express the model’s functional relation as Xt=f(Zt, β). Under specific assumptions about the structure of the economy (e.g. log or quadratic preferences, CobbDouglas production function, full depreciation of the capital stock in the one-sector growth model), f can be computed analytically either by value function iteration or by solving the Euler equations of the model subject to the transversality condition (see e.g. Hansen and Sargent, 1979). In general, however, f cannot be derived analytically from the primitives of the problem. A large body of current literature has concentrated on the problem of finding approximations which are either locally or globally close to f for a given metric.1 1 These include linear or log-linear expansions of f around the steady state (Kydland and Prescott, 1982; and King et al., 1988), backward-solving methods (Sims, 1984; Novales, 1990), global functional expansions in polynomials (Marcet, 1992; Judd, 1992), piecewise linear interpolation methods (Coleman, 1989; Baxter, 1991) and quadrature techniques (Tauchen and Hussey, 1991). © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 336 CALIBRATION S126 F.CANOVA Here we assume that either f is available analytically or that one of the existing numerical procedures has been employed to obtain a functional which approximates f in some sense, where γ=i(β) and is a given norm. Given the model f, i.e. an approximation procedure a set of parameters β, and a probability distribution for Zt (denoted by κ(Zt)), we can infer the model-based probability distribution for Xt. be the density of the Xt vector, conditional on the parameters β and the Let be the density of the parameters, conditional on the information set model f, let available to the simulator and the model f, and let be the joint density of represents the probability that a particular simulated data and of parameters. path for the endogenous variables will be drawn given a parametric structure for the summarizes the information on artificial economy and a set of parameters, while the parameters available to a researcher. Note that is assumed to be independent of and π may depend on f, i.e. if we are using a GE model we may want to use only estimates obtained with similar GE models. For a given β, Xt is random because Zt is random, i.e. is a deterministic transformation of k(Zt). Throughout this paper we are interested in studying the behaviour of functions of simulated data (denoted by µ(X t )) under the predictive density i.e. evaluating objects of the form: (1) where and is the parameter space and is the support of the exogenous variables. be the corresponding vector of functions of the actual data. Let The problem of measuring the fit of the model can be summarized as follows. How likely ? To answer note that from equation (1) we can compute is the model to generate probabilities of the form P(v(Xt)⑀D), where D is a bounded set and v(Xt) includes moments and other statistics of the simulated data. To do this let µ(Xt)=(Xt, [Xt:v(Xt) ⑀D]) where (Xt, S) is the indicator function, i.e. (Xt, S)=1 if v(Xt)⑀S and zero otherwise. Similarly, we can construct quantiles q(Xt) by appropriately choosing D (see e.g. Geweke, 1989). Finally, we satisfying for any given vector v, by appropriately can also find a selecting the indicator function. Model evaluation then consists of several types of statements which are interchangeable and differ only in the criteria used to measure distance. First, we can compute In other words, we can examine the likelihood of an event (the observed realization of the summary statistics in the actual data) from the point of view of the model. Extreme values for this probability indicate a poor ‘fit’ in the dimensions examined. we can then choose the set Alternatively, if we can measure the sampling variability of plus one or two standard deviations and either D to include the actual realization of check whether is in D or calculate P[v(Xt)⑀D]. Implementation There are four technical implementation issues which deserve some discussion. The first concerns the computation of integrals like those in equation (1). When the (β, Zt) vector is of high-dimension simple discrete grid approximations, spherical or quadrature rules quickly become infeasible since the number of function evaluations increases exponentially © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors STATISTICAL INFERENCE IN CALIBRATED MODELS STATISTICAL INFERENCE IN CALIBRATED MODELS 337 S127 with the dimension of β and Z t . In addition, unless the contours of are of ellipsoidal forms, grid evaluations may explore this density inefficiently. There are several feasible alternatives: one is the Monte Carlo procedure described in Geweke (1989), another is the data-augmentation procedure of Tanner and Wong (1987), a third is the ‘Gibbs sampler’ discussed in Gelfand and Smith (1990). Finally, we could use one of the quasi-random procedures proposed by Niederreiter (1988). In this paper we adopt a Monte Carlo approach. After drawing with replacement i.i.d. β vectors and Zt paths, we substitute sums over realizations for the integrals appearing in equation (1) and appeal to the strong law of large numbers for functions of i.i.d. random variables to obtain (2) where N is the number of replications. Note that, although (and p) are, in general, unknown, sampling from them can be conveniently accomplished by simulating the model repeatedly for random (Zt, β), i.e. randomly drawing exogenous forces and selecting a parameter vector and using the decision rule to compute time paths for Xt. Second, since in most cases the function f is unknown, itself becomes unknown and the direct computation of equation (1) is infeasible. If the approximation to f is accurate, we could simply neglect the error and proceed using in place of where is the joint density of simulated data and parameters using the information set and the approximation rule However, since only little is known about the properties of approximation procedures and some have only local validity (see e.g. Christiano, 1990; Dotsey and Mao, 1991), we may want to explicitly account for the existence of an approximation error in conducting inference. In this case, following Geweke (1989), we would replace equation (1) with: (3) are weights which depend on the ‘true’ density and on the where For example, if a quadratic approximation around approximation density the steady state is used, the density can be chosen so that draws of Zt inducing paths of Xt which are in the tails of (i.e. paths which are very far away from steady states) receive a very small weight in the calculation of the statistics of interest. Third, we must specify a density for the parameters of the model. We could select it on to be the asymptotic distribution of the basis of one specific data set and specify a GMM estimator (as in Burnside et al., 1993), of a simulated method of moments (SMM) estimator (as in Canova and Marrinan, 1993), or of a ML estimator of β (as in Phillips, 1991). The disadvantage of these approaches is that the resulting density measures the uncertainty surrounding β present in a particular data set and does not necessarily reflect the uncertainty faced by a researcher in choosing the parameters of the model. As Larry Christiano has pointed out to the author, once a researcher chooses the moments to match, the uncertainty surrounding estimates of β is small. The true uncertainty lies in the choice of moments to be matched and in the sources of data to be used to compute estimates. so as to summarize efficiently all existing A better approach would be to select information, which may include point estimates of β obtained from different estimation techniques, data sets, or model specifications. El-Gamal (1993a, b) has formally solved the © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 338 CALIBRATION S128 F.CANOVA problem of finding such a using Bayesian methods. The resulting is the least informative pseudo-posterior density on the parameter space which is consistent with a set of constraints describing the information contained in various estimation experiments. El-Gamal suggests a Gibbs sampler algorithm to compute this density but, in practice, there are simpler ways to construct empirical approximations to this type of density. One would be to count estimates of β previously obtained in the literature and construct by smoothing the resulting histogram. For example, if one of the elements of the β vector is the risk aversion parameter, counting estimates obtained over the last 15 years from fully specified general equilibrium models and smoothing the resulting histogram, we would obtain a truncated (below-zero) bell-shaped density, centred around two and very small massabove four. Alternatively, we could take what the profession regards as a reasonable range for β and assume more or less informative densities on the support depending on available estimates. If theoretical arguments suggest that the maximum range for e.g. the risk aversion parameter is [0, 20], we can put higher weights on the interval [1, 3] where most of the estimates lie. If for some parameters previous econometric evidence is scant, measurement is difficult, or there are no reasons to expect that one value is more likely than others, we could assume uniform densities on the chosen support. Available estimates of β are not necessarily independent (the same data set is used in many cases) and some are less reliable than others. Non-independent estimates are legitimate candidates to enter the information set as long as they reflect sampling variability or different estimation techniques. The influence of less reliable estimates or of estimates obtained with different model specifications can be discounted by giving them a smaller weight in constructing histograms (see also El-Gamal, 1993a, b). Finally, in many applications the joint density of the parameter vector can be factored into the product of lower-dimensional densities. If no relationship across estimates of the is the product of univariate densities. If estimates of various parameters exists, certain parameters are related (e.g. in the case of the discount factor and the risk aversion parameter in asset pricing models), we can choose multivariate densities for these dimensions and maintain univariate specifications for the densities of the other parameters. To summarize, to implement the procedure we need to do the following: • • • • • where represents the information Select a reasonable (data-based) density set available to a researcher, and a density κ(Zt) for the exogenous processes. and zt from κ(Zt). Draw vectors β from and compute µ(xt) using the model xt=f(zt, β) For each draw of β and zt, generate or the approximation Repeat the two previous steps N times. Construct the frequency distribution of µ(xt), compute probabilities, quantiles and other measures of interest. An Interpretation The proposed framework of analysis lends itself to a simple Bayesian interpretation. In this case we treat as the prior on the parameters. The function is entirely analogous to a classical likelihood function for Xt in a standard regression model. The difference is that need not be the correct likelihood for and need not have a closed form. Then equation (1) is the conditional expectation of µ(Xt) under the predictive density of the model and probability statements based on equation (1) can be justified using the arguments contained in Box (1980). © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors STATISTICAL INFERENCE IN CALIBRATED MODELS STATISTICAL INFERENCE IN CALIBRATED MODELS 339 S129 There is also a less orthodox interpretation of the approach which exchanges the role of and and is nevertheless reasonable. In this case is the prior and represents the a priori degree of confidence posed by the researcher on the time summarizes the path generated by the model given the parameters while information contained in the data. Then equation (1) is a ‘pseudo-posterior’ statement about the model’s validity once the empirical evidence on the parameters is taken into account. It is useful to note that, if we follow the first approach, we can relate the proposed to the data-based priors employed in Meta-Analysis (see Wolf, construction of 1986) and in the ‘consensus literature’ (see e.g. Genest and Zideck, 1986). El-Gamal (1993a) spells out in detail the connection with these two strands of literature. 3. ROBUSTNESS ANALYSIS If we adopt a Monte Carlo approach to compute simulated densities for the statistics of interest, an automatic and efficient global sensitivity analysis is performed on the support of the parameter space as a by-product of the simulations. Sensitivity analysis, however, can take other more specific forms. For example, we may be interested in examining how likely when β is fixed at some prespecified value This would be the µ(Xt) is to be close to case, for example, if β includes parameters which can be controlled by the government and is e.g. the current account balance of that country. In this case we could choose a path for Zt and analyse the conditional distribution of µ(Xt) for the selected value(s) of β. Alternatively, we might wish to assess the maximal variation in µ(Xt) consistent, say, with β being within two standard deviations of a particular value. Here we choose a path for Zt and construct paths for µ(Xt) for draws of β in the range. Finally, we may be interested in knowing which dimensions of β are responsible for particular features of the distribution of µ(Xt). For example, if the simulated distribution of µ(Xt) has a large spread or fat tails, a researcher may be interested in knowing whether technology or preference parameters are responsible for this feature. In this case we would partition β into [β1, β2] and compute the where is a prespecified value (or simulated distribution of µ(Xt) conditional on set of values. So far, we have examined the robustness of the results to variations of the parameters within their support. In some cases it is necessary to study the sensitivity of the results to local perturbations of the parameters. For example, we may be interested in determining how robust the simulation results are to changes of the parameters in a small neighbourhood of a particular vector of calibrated parameters. To undertake this type of analysis we can take a numerical version of the average derivative of µ(X t) in the neighbourhood of a calibrated vector (see Pagan and Ullah, 1991). Because global and local analyses aim at examining the sensitivity of the outcomes to perturbations in the parameters of different size they provide complementary information and should both be used as specification diagnostics for models whose parameters are calibrated. 4. A COMPARISON WITH EXISTING METHODOLOGIES The framework of analysis in Section 2 is general enough to include simulation undertaken after the parameters are both calibrated and estimated via method of moments as special is a deterministic transforcases. To show this it is convenient to recall that The two procedures can then be recovered by mation of and, in some cases, also on imposing restrictions on the shape and the location of the shape and the location of k(Zt). © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 340 S130 CALIBRATION F.CANOVA Calibration exercises impose a point mass for on a particular value of β (say, ). One interpretation of this density selection is that a simulator is perfectly confident in the vector β used and does not worry about the cross-study or time-series uncertainty surrounding estimates of β. In certain situations a path for the vector of exogenous variables is also selected in advance either by drawing only one realization from their distribution or by choosing a zt on the basis of extraneous information, e.g. inputting Solow’s residuals in the model, so that κ(Zt) is also a singleton. In this instance, the density of µ(Xt) has a point mass and because the likelihood of the model to produce any event is either 0 or 1, we must resort to informal techniques to assess the discrepancy of simulated and actual data. In some studies the randomness in Zt is explicitly considered, repeated draws for the exogenous variables are made for a fixed and moments of the statistics of interest are computed by averaging the results over a number of simulations (see e.g. Backus et al., 1989). Simulation exercises undertaken with estimation of the parameters are also special cases has a point mass at β*, where β* is either the GMM of the above framework. Here estimator, the SMM estimator (see Lee and Ingram, 1991), or the simulated quasimaximum likelihood (SQML) estimator of β (see Smith, 1993). Simulations are typically and standard errors for µ(Xt) performed by drawing one realization from are computed using the asymptotic standard errors of β* and the functional form for µ. In some cases, is taken to be the asymptotic distribution of one of the above estimators (e.g. Canova and Marrinan, 1993). In this case, simulations are performed by drawing from and the distance of simulated and actual data is computed using measures of discrepancy like the ones proposed here. In assessing the model’s performance this last set of procedures has two advantages over calibration. First, they allow formal statements on the likelihood of selected parameter values to reproduce the features of interest. For example, if a four standard deviations range around the point estimate of the AR(1) parameter for the productivity disturbance is [0·84, 0·92], then it is highly unlikely (with a probability higher than 99%) that a unit root productivity disturbance is needed to match the data. Second, they provide a set-up where sensitivity analysis can easily be undertaken (although not often performed). These procedures, however, have also two major shortcomings. First, they impose a strong form of ignorance which does not reflect available a priori information. The vector β may include meaningful economic parameters which can be bounded on the basis of theoretical arguments but the range of possible β with GMM, SMM, or SQML procedures is [-∞, ∞]. By appropriately selecting a hypercube for their densities a researcher can make ‘unreasonable’ parameter values unlikely and avoid a posteriori adjustments. Second, simulations conducted after parameters are estimated may not constitute an independent way to validate the model because the parameter estimates are obtained from the same data set which is used later to compare results. Simulation procedures where parameters are selected using a mixture of calibration and estimation strategies have recently been employed by e.g. Heaton (1993) and Burnside et al. (1993). Here some parameters are fixed using extraneous information while others are formally estimated using moment (or simulated moment) conditions. Although these strategies allow a more formal evaluation of the properties of the model than pure calibration procedures, they face two problems. First, as in the case when the parameters are all selected using GMM, SMM, and SQML procedures, the evaluation of the model is problematic because measures of dispersion for the statistics of interest are based on one data set and do not reflect the uncertainty faced by a simulator in choosing the unknown features of the model. Second, as Gregory and Smith (1989) have pointed out, the smallsample properties of estimators obtained from these strategies may be far from reasonable © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors STATISTICAL INFERENCE IN CALIBRATED MODELS STATISTICAL INFERENCE IN CALIBRATED MODELS 341 S131 unless calibrated parameters consistently estimate the true parameters. When this condition is not satisfied, estimates of the remaining parameters are sensitive to errors in pre-setting and results are misleading. The Monte Carlo methodology we employ to evaluate the properties of the model is related to those of Kwan (1990) and Gregory and Smith (1991) but several differences need to be emphasized. First, Gregory and Smith take the model as a testable null hypothesis while this is not the case here. Second, they do not account for parameter uncertainty in evaluating the outcomes of the model. Third, because they take a calibrated version of the model as the ‘truth’, they conduct sensitivity analysis inefficiently, by replicating experiments for different calibrated values of the parameters. Kwan (1990) allows for parameter uncertainty in his simulation scheme, but, following an orthodox Bayesian approach, he chooses a subjective prior density for the parameters. In addition, he evaluates the outcomes of the model in relative terms by comparing two alternative specifications using a posterior-odds ratio: a model is preferred to another if it maximizes the probability that the simulated statistics are in a prespecified set. The procedure for sensitivity analysis we proposed extends the approach that Harrison and Vinod (1989) used in deterministic computable general equilibrium models. The major difference is that in a stochastic framework parameter uncertainty is only a part of the randomness entering the model and the uncertainty characterizing the exogenous processes is important in determining the randomness of the outcomes. To conclude, we should mention that, parallel to the literature employing Monte Carlo methods to evaluate calibrated models, there is also a branch of the literature which uses alternative tools to examine the fit of calibrated models. This is the case e.g. of Smith (1993), Watson (1993), and Canova et al. (1993) which assess the relevance of theoretical models with regression R2’s, tests based on restricted and unrestricted VARs, and encompassing procedures. 5. TWO EXAMPLES The Equity Premium Puzzle Mehra and Prescott (1985) suggest that an asset-pricing model featuring complete markets and pure exchange cannot simultaneously account for the average risk-free rate and the average equity premium experienced by the US economy over the sample 1889–1978 with reasonable values of the risk aversion parameter and of the discount factor. The model they consider is a frictionless Arrow-Debreu economy with a single representative agent, one perishable consumption good produced by a single productive unit or a ‘tree’, and two assets, an equity share and a risk-free asset. The tree yields a random dividend each period and the equity share entitles its owner to that dividend in perpetuity. The risk-free asset entitles its owner to one unit of the consumption good in the next period only. The agent maximizes: (4) subject to: (5) where E0 is the mathematical expectation operator conditional on information at time zero, is the subjective discount factor, is the risk aversion parameter, ct is consumption, yt is the © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 342 S132 CALIBRATION F.CANOVA tree’s dividend, and are the prices of the equity and the risk-free asset, and et and ft are the agent’s equity and risk-free asset holding at time t, Production evolves according to yt+1=xt+1yt where xt, the gross growth rate, follows a two-state ergodic Markov chain with Defining the states of the problem as (c, i) where yt=c probability and xt=λi, the period t equilibrium asset prices are (6) (7) When the current state is (c, i), the expected equity return and the risk-free rate are: (8) (9) The unconditional (average) expected returns on the two assets are and the average equity premium is EP=Re-Rf, where πi are the Markov chain and Σiπi=1, where stationary probabilities, satisfying Mehra and Prescott specified the two states for consumption (output) to be λ1=1+µ+v; λ2=1+µ-v and restricted and They calibrated the three technology parameters so that the mean, the standard deviation, and the AR(1) coefficient of the model’s consumption match those of the growth rate of annual US consumption over the period 1889–1978 and searched for combinations of the preference parameters (, ) in a prespecified interval to obtain values for the risk-free rate and the equity premium. Given that the average, the standard deviations, and the AR(1) coefficient of annual growth rate of US consumption are 1·018, 0·036, and -0·14, the implied values of µ, v, and are 0·018, 0·036, and 0·43, respectively. The range for was selected to be [0, 10] and this choice was justified citing a number of empirical estimates of this parameter (see Mehra and Prescott, 1985, p. 154). The range for ω was chosen to be [0, 1], but simulations which produced a risk-free rate in excess of 4% were eliminated on the grounds that 4% constitutes an upper bound consistent with historical experience. The puzzle is that the largest equity premium generated by the model is 0·35%, which occurred in conjunction with a real risk-free rate of about 4%, while the US economy for the period 1889–1978 experienced an annual average equity premium of 6·18% and an average real risk-free rate of 0·80%. Two hidden assumptions underlie Mehra and Prescott’s procedure. First, they believe that technology parameters can be tightly estimated while the uncertainty surrounding the choice of preference parameters is substantial. Consequently, while the sensitivity of the results is explored to variations in θ and ω within the range, no robustness check is made for perturbations of the technology parameters. Second, by providing only the largest value generated, they believe that it is a sufficient statistic to characterize the puzzle. Here we repeat their exercise with three tasks in mind: first, to study whether the uncertainty present in the selection of the technology parameters is important in determining the magnitude of the puzzle; second, to formally measure the discrepancy of © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors STATISTICAL INFERENCE IN CALIBRATED MODELS STATISTICAL INFERENCE IN CALIBRATED MODELS 343 S133 the model from the data using a variety of statistics based on the probability distribution of outcomes of the model; third, to evaluate the contribution of two alleged solutions to the equity premium puzzle proposed in the literature. This first example is particularly simple since we have an exact solution for the endogenous variables of the model. In addition, because the model produces values for the mean of Rf and EP, variations in Xt are entirely determined by variations in β, so that is proportional to Therefore, once we have selected we can immediately determine the distribution of simulated means of the Rf-EP pair. To select the density for the five parameters of the model we proceed in two steps. First, we choose a maximum range for the support of β on the basis of theoretical considerations. Table I. Equity premium puzzle Note: Pr 1 refers to the frequency of simulations for which the pair (Rf, EP) is in a classical 95% region around the actual values. Pr 2 reports the percentile of the simulated distribution where the actual (Rf, EP) pair lies. Pr 3 reports the probability that the model generates values in each of the four quadrants delimited by the actual pair of (Rf, EP). Q1 is the region where Rf1>Rf and EPs<EP, Q2 is the region where Rfs>Rf and EP4ⱖEP, Q3 is the region where RfsⱕRf and EP4<EP and Q4 is the region where RfsⱕRf and EPsⱖEP. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 344 S134 CALIBRATION F.CANOVA Second, we specify the joint density to be the product of five univariate densities and select each univariate density to be a smoothed version of the frequency distribution of estimates existing in the literature. The densities and their support are in panel A of Table I. Therange for ω is the same as that of Mehra and Prescott and the chosen χ2 density has a mode at 2, where most of the estimates of this parameter lie, and a low mass (smaller than 5%) for values exceeding 6. The range for θ reflects the results of several estimation studies which obtained values for the steady-state real interest rate in the range [-0·005, 0·04] (see e.g. Altug, 1989; Dunn and Singleton, 1986; or Hansen and Singleton, 1983) and of simulation exercises which have a steady-state real interest rate in the range [0, 0·05] (see e.g. Kandel and Stambaugh, 1990; or Mehra and Prescott, 1985). The density for is skewed to express the idea that a steady-state real interest rate of 2–3% or lower is more likely than a steadystate interest rate in excess of 4%. Note that although we assume that the densities of θ and ω are independent, many estimates of these two parameters are not. However, the rank correlation coefficient for the pairs of estimates is small and none of the results we present depends on this simplifying assumption. To provide a density for µ, v and we experimented with two procedures. The first, which is used in the basic experiment, involves taking the 10 sub-sample estimates of the mean, of the standard deviation, and of the AR(1) coefficient of the growth rate of consumption over 10-year samples contained in Mehra and Prescott (1985, p. 147) as characterizing reasonable consumption processes and then constructing a uniform discrete density over these triplets. The second involves dividing the growth rates of consumption over the 89 years of the sample into two states (above and below the mean), estimating a measure of dispersion for the first two moments and for the AR(1) coefficient of the growth rate of consumption in each state and directly inputting these estimates into the model. In this case simulations are performed by assuming a joint normal density for the mean, the standard deviation, and AR(1) coefficient in each state centred around the point estimate of the parameters and maximum support within two standard deviations of the estimate. Figures 1–4 present scatterplots of the simulated pairs (Rf, EP) when 10,000 simulations are performed. We summarize the features of the joint distribution in panel B of Table I Figure 1. Scatterplot risk-free rate-equity premium: Mehra-Prescott case © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors STATISTICAL INFERENCE IN CALIBRATED MODELS STATISTICAL INFERENCE IN CALIBRATED MODELS Figure 2. Scatterplot risk-free rate-equity premium: basic case Figure 3. Scatterplot risk-free rate-equity premium: beta>1 case © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 345 S135 346 S136 CALIBRATION F.CANOVA Figure 4. Scatterplot risk-free rate-equity premium: Reitz case using a number of statistics. To evaluate the discrepancy of the model from the data we report (1) the probability that the model generates values for (Rf, EP) which fall within a two standard deviation band of the actual mean, (2) the percentile contour of the simulated distribution where the actual means of (Rf, EP) lies, and (3) the probability that the simulated pair is in each of the four quadrants of the space delimited by the actual means of (Rf, EP). Figure 1 reports the scatterplot obtained with the Mehra and Prescott specification (i.e. when technology parameters are fixed and we draw replications from the densities of θ and ω only). It is necessary to check that the maximum value of the equity premium consistent with a risk free-rate not exceeding 4% is only 0·0030, confirming Mehra and Prescott’s conclusion. Also for this specification, the distribution of the model’s outcomes is uniform and the mode of the joint distribution (the most likely value from the point of view of the model) is at Rf=0·110, EP=0·0094. The probabilistic measures of discrepancy suggest that a large portion of the simulations are in the region where the simulated Rf exceeds the mean of Rf and the simulated EP is below the mean of EP we find in the data, that about 73% of the simulations produce pairs within a classical 95% ball around the actual means of (Rf, EP), and that the actual mean pair is outside the 99 percentile contour. Figure 2 reports the scatterplot obtained with the basic specification of the model. Also in this case, the puzzle, as defined by Mehra and Prescott, is evident: if we set a 4% upper bound to the risk-free rate, the maximum equity premium generated is only 0·0038. However, with this specification, the distribution is bimodal and most of the simulated pairs lie on a ridge parallel to the Rf axis. The probability that the model generates values in a ball centred around the actual means of (Rf, EP) is now 81·4%. However, in 100% of the cases the simulated risk-free rate exceeds the actual mean and the simulated equity premium is below the actual mean and the actual pair still lies outside the 99 percentile contour of simulated distribution. To examine whether the selection of the density for the technology parameters has effects on the results, we also conducted simulations using the alternative distribution for these parameters. No substantial changes emerge. For example, the probability that the © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors STATISTICAL INFERENCE IN CALIBRATED MODELS STATISTICAL INFERENCE IN CALIBRATED MODELS 347 S137 model generates pairs in a ball centred around the actual means of (Rf, EP) is 80·3% and the maximum value for EP compatible with a Rf not exceeding 4% is 0·0025. Several conclusions can be drawn from this first set of exercises. First, even after taking into account the uncertainty surrounding estimates of the technology parameters, the puzzle remains regardless of the way it is defined (maximum values, modes, or contour probabilities): the model cannot generate (Rf, EP) pairs which match what we see in the data. Second, once the uncertainty surrounding estimates of the technology parameters is taken into account, the simulated distributions are bimodal, highly left skewed, and have a fat left tail, indicating that lower than average values are more probable and that very small values have nonnegligible probability. Third, the simulated risk-free rate is always in excess of the actual one, a result that Weil (1990) has termed the risk-free rate puzzle. Fourth, while the model fails to generate values for (Rf, EP) which replicate the historical experience, in more than 80% of the simulations it produces pairs which are within two standard deviations of the actual means. Next, we conduct two exercises designed to examine the contribution of the modifications suggested by Kocherlakota (1990), Benninga and Protopapadakis (1990), and Rietz (1988) to the solution of the puzzle. The first experiment allows the discount factor θ to take on values greater than 1. The justification is that, in a growing economy, reasonable values for the steady-state real interest rate can be obtained even with θ greater than 1. In this experiment we still maintain the truncated normal density for θ used in the baseline case but increase the upper value for its range to 1·04 and allow about 10% of the density in the region above 1.0. The second experiment assumes the presence of a third unlikely crash state where consumption falls substantially.2 The justification for including a third state is that in the Great Depression consumption fell substantially and excluding such a state may have important implications on the results (a conclusion denied by Mehra and Prescott, 1988). With this specification there are two new parameters which cannot be measured from available data: ξ, the probability of a crash state and ξ, the percentage fall in consumption in the crash state. Rietz (1988) searched over the a priori ranges of [0·0001, 0·2] and [µ/(1+µ), 1v/(1+µ)] and examined the magnitude of the maximum simulated equity premium that the model consistent with a simulated risk-free rate below 4%. We maintain these ranges in our experiment and assume on these supports an exponential density for ξ and a three-point discrete density for ξ summarizing the three cases examined by Rietz. Allowing the discount factor to take on values greater than 1 goes a long way towards reducing the discrepancy of the model from the data (see Figure 3) since it shifts the univariate distribution of Rf towards negative values (the minimum and maximum values of Rf are now (-0·084, 0·0.092). For example, the probability that the model generates pairs in a ball centred around the actual means of (Rf, EP) is now 85·7% and in only 7·4% of the cases is the simulated risk-free rate in excess of the actual means. Because of this shift in the univariate distribution of Rf, the maximum value of EP consistent with a risk-free rate below 4% is now 0·031. Despite these differences, the location and the shape of the univariate distribution of EP are unaffected. Hence, although the equity premium puzzle is ‘solved’ when defined in terms of the maximum simulated EP consistent with a simulated Rf below 4%, it is still very evident when we look at the distributional properties of the simulated EP. 2 The three consumption states are and the transition matrix has elements: Note that the experiment is conceptually different from the previous ones since there are two extra degrees of freedom (the new parameters ξand ξ) and no extra moments to be matched. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 348 CALIBRATION S138 F.CANOVA The second modification is much less successful (see Figure 4). It does shift the univariate distribution of EP to the right (the mode of 0·035) and increases the dispersion of simulated EPs but it achieves this at the cost of shifting the distribution of Rf towards unrealistic negative value (the mean is -0·15 and the 90% range is [-0·940, 0·068]) and of scattering the simulated (Rf, EP) pairs all over the place. For example, the probability that the simulated pair is in a ball centred around the actual means of (Rf, EP) decreases to 72·7% and the probabilities that the model generates values in each of thefour quadrants delimited by the actual means of (Rf, EP) are almost identical. Finally, the maximum EP consistent with a Rf below 4% is 0·747. Therefore, adding a crash state shifts the mode and stretches and tilts the shape of the joint simulated distribution. Roughly speaking, too many (Rf, EP) configurations now have equal probability, and this weakens the ability of the theory to provide a coherent answer to the question posed. Technology Shocks and Cyclical Fluctuations in GNP Kydland and Prescott (1982) showed that a one-sector growth model driven by technology shocks calibrated to reproduce the statistical properties of Solow residuals explains about 70% of the variance of per capita US output. This result has spurred much of the subsequent literature which tries to account for business cycle regularities in models where monetary impulses play no role (the so-called real business cycle literature). Kydland and Prescott’s initial estimate has been refined by adding and subtracting features to the basic model (see Hansen, 1985) but the message of their experiment remains: a model where technology shocks are the only source of disturbance explains a large portion of the variability of per capita US output. Recently, Eichenbaum (1991) has questioned this assertion because ‘decisions based and and solely on the point estimate of λy are whimsical var(yt) are the variance of the cyclical component of simulated and actual output) and suggests that ‘the model and the data, taken together, are almost completely uninformative about the role of technology shocks in generating fluctuations in US output’ (pp. 614–615). Using an exactly identified GMM procedure to estimate the free parameters, he finds that the model explains anywhere between 5% and 200% of the variance of per capita US output. In this section we repeated Eichenbaum’s exercise with three goals in mind. First, we are interested in knowing that is the most likely value of λy from the point of view of the model (i.e. in locating the mode of the simulated distribution). Second, we want to provide confidence bands fo λy which reflect the uncertainty faced by a researcher in choosing the parameters of the model (not the uncertainty present in the data, as in Eichenbaum). Third, we wish to verify whether normal confidence bands appropriately describes the uncertainty surrounding point estimates of λy and examine which feature of the model make deviations from normality more evident. The model is the same as Eichenbaum’s and is a simple variation of Hansen’s (1985) model which allows for deterministic growth via labour-augmenting technological progress. The social planner of this economy maximizes (10) subject to: (11) © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors STATISTICAL INFERENCE IN CALIBRATED MODELS STATISTICAL INFERENCE IN CALIBRATED MODELS 349 S139 where ct is per capital consumption, T-ht is leisure, and kt the capital stock. When δ is different from 1, a closed-form stationary solution to the problem does not exist. Here we compute an approximate decision rule for the endogenous variables using a loglinear expansion around the steady state after variables have been linearly detrended as in King et al. (1988), but we neglect the approximation error in constructing probability statements on and no weighting). the outcomes of the model (i.e. we use There are seven parameters in the model, five deep (δ, the depreciation rate of capital; β, the subjective discount rate; , leisure’s weight in the utility function; α, labour’s share in output; γ, the constant unconditional growth rate of technology) and two which appear only because of the auxiliary assumptions we made on the stochastic process for technology shocks (ρ, the AR parameter and σ the standard deviation of the shock). Hansen (1985) calibrated these seven parameters (the values are in the first column of panel A of Table II) and found that λy≈1. Eichenbaum (1991) estimated all parameters except β (which is calibrated) using a method of moments estimator (estimates and standard deviations are in the second column of panel A of Table II) and found (1) a point estimate of λy of 0·80, (2) a large standard deviation about the point estimate of λy due primarily to the uncertainty surrounding estimates of ρ and σ, and (3) a strong sensitivity of the point estimate of λy to small perturbations in the parameter vector used. Table II. Technology shocks and cyclical fluctuations in GNP Note: Estimated standard errors are in parentheses. Pr 1 refers to the frequency of simulations for which the variance of simulated output is in a classical 95% region around the actual value of the variance of detrended output. Pr 2 reports the percentile of the simulated distribution where the point estimate of the actual variance of output lies. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 350 S140 CALIBRATION F.CANOVA In the exercise we conduct, we assume that is the product of seven univariate densities. Their specification appear in the third column of panel A of Table II The range for the quarterly discount factor corresponds to the one implied by the annual range used in the previous example and the density is the same. ␦ is chosen so that the annual depreciation rate of the capital stock is uniformly distributed between 8% and 12% per year. The range is selected because in simulation studies ␦ is commonly set to 0·025, which corresponds to a 10% annual depreciation rate, while estimates of this parameter lie around this value (e.g. McGratten et al., 1991, have a quarterly value of 0·0310 and a standard deviation of 0·0046, while Burnside et al., 1993, have a quarterly value of 0·0209 and a standard deviation of 0·0003). The range for a reflects calculations appearing in Christiano (1988) where, depending on how proprietors’ income is treated, the share of total output paid to capital varies between 0·25 and 0·43, and the estimate obtained, among others, in McGratten et al. (1991). We chose the densities for ρ and σ as in Eichenbaum because the econometric evidence on these two parameters is scant andthe values used in most simulation studies fall within a one standard deviation band around the mean of the assumed density (see e.g. Kydland and Prescott, 1982; Hansen, 1985). Finally, T is fixed at 1369 hours per quarter, the density for γ matches the quarterly distribution of unconditional quarterly growth rates of US output for the period 1950–1990, and is endogenously chosen so that the representative household spends between one sixth and one third of its time working in the steady state. We performed 1000 simulations with time series of length T=124 and filtered both simulated and actual GNP data with the Hodrick and Prescott filter.3 The results appear in panel B of Table II and in Figure 5, where we present a smoothed version of the simulated distribution of λy. The distribution is scaled so that with the point estimates of the parameters used by Eichenbaum λy=0·80. The implied value of λy using Hansen’s parameters is 0·84. Figure 5. Density of variance ratio: HP filtered data 3 We use the Hodrick and Prescott filter to maintain comparability with previous work. The results obtained when the data are linearly detrended or first-order differenced are not substantially different. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors STATISTICAL INFERENCE IN CALIBRATED MODELS STATISTICAL INFERENCE IN CALIBRATED MODELS 351 S141 The mode of the distribution of λy is at 0·9046, the mean at 0·8775, and the median at 0·5949. The dispersion around these measures of location is very large. For example, the standard deviation is 0·7635 and the 90% range of the distribution is [0·2261, 2·6018]. The simulated distribution is far from normal and its right tail tends to be very long. Hence the range of reasonable values of λy is very large, and, as in Eichenbaum, small perturbations in the parameter vector induce large variations in the variance ratio. In addition, normal confidence bands do not appropriately characterize the uncertainty surrounding the outcomes of the model. Several other features of the simulated distribution are worth mentioning. First, in 67·3% of the cases the variance of simulated output is smaller than the variance of actual output. Second, in 42·7% of the simulations the variance of simulated output is within a 95% confidence interval centred around the estimate of the variance of actual output. i.e. is the median Third, if we select v=0·5 and look for the satisfying of the simulated distribution, we find that the median value of the variance of simulated GNP is outside the 95% normal confidence interval for the variance of actual GNP. When we ask which parameter is responsible for the wide dispersion in the estimates of λy, we find that it is the location and width of the support of ρ which induce this feature in the distribution of λy. For example, assuming that the density of ρ has a point mass at 0·94 and maintaining the same densities for the other parameters, we find that location measures of the simulated distribution of λy decrease (the mode is now at 0·792) and the standard deviation drops to 0·529. Similar conclusions are obtained by shifting the range of ρ towards 0·90 or by cutting the range of possible ρ in half without changing the mean value. Hence, as in Eichenbaum, we find that it is the uncertainty present in the choice of the parameters of the exogenous processes rather than the uncertainty present in the selection of the deep parameters of the model that is responsible for the large spread in the distribution of λy. 6. CONCLUSIONS This paper describes a Monte Carlo procedure to evaluate the properties of calibrated general equilibrium models. The procedure formalizes the choice of the parameters and the evaluation of the properties of the model while maintaining the basic approach used in calibration exercises. It also realistically accounts for the uncertainty faced by a simulator in choosing the parameters of the model. The methodology allows for global sensitivity analysis for parameters chosen within the range of existing estimates and evaluates the discrepancy of the model from the data by attaching probabilities to events a simulator is interested in characterizing. The approach is easy to implement and includes calibration and simulation exercises conducted after the parameters are estimated by simulation and GMM techniques as special cases. We illustrate the usefulness of the approach as a tool to evaluate the performance of theoretical models with two examples which have received much attention in the recent macroeconomic literature: the equity premium puzzle and the ability of a real business cycle model to reproduce the variance of actual US output. Finally, it is worth noting that for problems of moderate size, the computational complexity of the procedure is limited. For both examples presented the entire Monte Carlo routine required about a minute on a 486–33 MHz machine using RATS386 programs. ACKNOWLEDGEMENTS Part of this research was undertaken while the author was also associated with the European University Institute, Florence. The author has benefited from the comments and the © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 352 S142 CALIBRATION F.CANOVA suggestions of a large number of colleagues, including two anonymous referees, David Backus, Larry Christiano, Frank Diebold, Javier Diaz, Mahmoud El-Gamal, John Geweke, Eric Ghysels, Bruce E.Hansen, Gary Hansen, Jane Marrinan, Yaw Nyarko, Adrian Pagan, Franco Peracchi, Victor Rios, Gregor Smith, Herman Van Dijk, and Randall Wright. He would also like to thank the participants of seminars at Brown, European University Institute, NYU, Montreal, Rochester, Penn, University of Rome, Carlos III Madrid, Free University Bruxelles, CERGE, University of Minnesota, University of Maryland, Summer Meetings of the North American Econometric Society and the Conference on ‘Econometric Inference Using Simulation Techniques’ held in Rotterdam on 5–6 June 1992 for useful discussions. REFERENCES Altug, S. (1989), ?‘Time to build and aggregate fluctuations: some new evidence’, International Economic Review, 30, 889–920. Backus, D., A. Gregory and A.Zin (1989), ‘Risk premiums in the terms structure: evidence from artificial economies’, Journal of Monetary Economics, 24, 371–399. Baxter, M. (1991), ‘Approximating suboptimal dynamic equilibria: an Euler equation approach’, Journal of Monetary Economics, 27, 173–200. Benninga, S. and A. Protopapadakis (1990), ‘Leverage, time preference, and the equity premium puzzle’, Journal of Monetary Economics, 25, 49–58. Box, G. (1980), ‘Sampling and Bayes inference in scientific modelling and robustness’, Journal of the Royal Statistical Society, Ser. A, 143, 383–430. Burnside, C., M.Eichenbaum and S.Rebelo (1993), ‘Labor hoarding and the business cycle’, Journal of Political Economy, 101, 245–273. Canova, F. and J.Marrinan (1993), ‘Profits, risk and uncertainty in foreign exchange markets’, Journal of Monetary Economics, 32, 259–286. Canova, F., M. Finn and A.Pagan (1993), ‘Evaluating a real business cycle model’, forthcoming, in C. Hargreaves (ed.), Nonstationary Time Series Analyses and Cointegration, Oxford: Oxford University Press. Christiano, L. (1988), ‘Why does inventory investment fluctuate so much?’ Journal of Monetary Economics, 21, 247–280. Christiano, L. (1990), ‘Solving the stochastic growth model by linear quadratic approximation and by value function iteration’, Journal of Business and Economic Statistics, 8, 23–26. Coleman, W. (1989), ‘An algorithm to solve dynamic models’, Board of Governors of the Federal Reserve System, International Finance Division, Discussion Paper No. 351. Dotsey, M. and C.S.Mao (1991), ‘How well do linear approximation methods work? Results for suboptimal dynamic equilibria’, Journal of Monetary Economics, 29, 25–58. Dunn, D. and K.Singleton (1986), ‘Modelling the term structure of interest rates under non-separable utility and durability of goods’, Journal of Financial Economics, 17, 27–55. Eichenbaum, M. (1991), ‘Real business cycle theory: wisdom or whimsy?’ Journal of Economic Dynamic and Control, 15, 607–621. El-Gamal, M. (1993a), ‘The extraction of information from multiple point estimates’, forthcoming in Journal of Nonparametric Statistics. El-Gamal, M. (1993b), ‘A Bayesian interpretation of extremum estimators’, manuscript, California Institute of Technology. Fair, R. (1991), ‘Estimating event probabilities from macroeconomic models using stochastic simulations’, Yale University, manuscript. Friedman, M. (1959), Essays in Positive Economics, New York: Aldine Press. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors STATISTICAL INFERENCE IN CALIBRATED MODELS STATISTICAL INFERENCE IN CALIBRATED MODELS 353 S143 Frisch, R. (1933), ‘Propagation problems and impulse problems in dynamics economies’, in Economic Essays in Honor of Gustav Cassel, London. Geweke, J. (1989), ‘Bayesian inference in econometric models using Monte Carlo integration’, Econometrica, 57, 1317–1339. Gelfand, A. and A.Smith (1990), ‘Sampling based approaches to calculating marginal densities’, Journal of the American Statistical Association, 85, 398–409. Genest, C. and M.Zidak (1986), ‘Combining probability distributions: a critique and an annotated bibliography’, Statistical Science, 1, 114–148. Gregory, A. and G.Smith (1989), ‘Calibration as estimation’, Econometric Reviews, 9(1), 57– 89. Gregory, A. and G.Smith (1991), ‘Calibration as testing: inference in simulated macro models’, Journal of Business and Economic Statistics, 9(3), 293–303. Haavelmo, G. (1944), ‘The probability approach in econometrics’, Econometrica, 12, Supplement. Hansen, G. (1985), ‘Indivisible labor and the business cycle’, Journal of Monetary Economics, 16, 309–328. Hansen, L. and T.Sargent (1979), ‘Formulating and estimating dynamic linear rational expectations models’, Journal of Economic Dynamic and Control, 2, 7–46. Hansen, L. and K.Singleton (1983), ‘Stochastic consumption, risk aversion and temporal behavior of asset returns’, Journal of Political Economy, 91, 249–265. Harrison, G. and H.D.Vinod (1989), ‘Sensitivity analysis of applied general equilibrium models: completely randomized factorial sampling designs’, University of New Mexico, manuscript. Heaton, J. (1993), ‘The interaction between time nonseparable preferences and time aggregation’, Econometrica, 61, 353–381. Journal of Business and Economic Statistics, January 1990. Judd, K. (1992), ‘Projection methods for solving aggregate growth models’, Journal of Economic Theory, 58, 410–452. Kandel, S. and R.Stambaugh (1990), ‘Expectations and volatility of consumption and asset returns’, Review of Financial Studies, 3, 207–232. King, R., C.Plosser and S.Rebelo (1988), ‘Production, growth and business cycles: I and II’, Journal of Monetary Economics, 21, 195–232 and 309–342. Kocherlakota, N. (1990), ‘On the discount factor in growth economies’, Journal of Monetary Economics, 25, 45–48. Kydland, F. (1992), ‘On the econometrics of world business cycles’, European Economic Review, 36, 476–482. Kydland, F. and E.Prescott (1982), ‘Time to build and aggregate fluctuations’, Econometrica, 50, 1345–1370. Kydland, F. and E.Prescott (1991), ‘The econometrics of the general equilibrium approach to business cycles’, The Scandinavian Journal of Economics, 93(2), 161–178. Kwan, Y.K. (1990), ‘Bayesian calibration with an application to a non-linear rational expectation two country model’, mimeo, University of Chicago Business School. Lee, B.S. and B.Ingram (1991), ‘Simulation estimators of time series models’, Journal of Econometrics, 47(2/3), 197–206. Marcet, A. (1992), ‘Solving nonlinear stochastic models by parametrizing expectations: an application to asset pricing with production’, Universitat Pompeu Fabra, Working Paper 5. Mehra, R. and E.Prescott (1985), ‘The equity premium: a puzzle’, Journal of Monetary Economics, 15, 145–162. Mehra, R. and E.Prescott (1988), ‘The equity risk premium: a solution?’ Journal of Monetary Economics, 22, 133–136. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors 354 S144 CALIBRATION F.CANOVA McGratten, E., R.Rogerson and R.Wright (1991), ‘Estimating the stochastic growth model with household production’, Federal Reserve Bank of Minneapolis, manuscript. Niederreiter, H. (1988), ‘Quasi Monte Carlo methods for multidimensional numerical integration’, International Series of Numerical Mathematics, 85, 157–171. Novales, A. (1990), ‘Solving nonlinear rational expectations models: a stochastic equilibrium model of interest rates’, Econometrica, 58, 93–111. Pagan, A. and A.Ullah (1991), ‘Nonparametric estimation’, University of Rochester, manuscript. Phillips, P.C.B. (1991), ‘To criticize the critics: an objective Bayesian analysis of stochastic trends’, Journal of Applied Econometrics, 6(4), 333–354. Prescott, E. (1991), ‘Real business cycle theories: what have we learned?’ Federal Research of Minneapolis, Working Paper 486. Reitz, T. (1988), ‘The equity risk premium: a solution’, Journal of Monetary Economics, 22, 117–132. Shoven, J. and J.Whalley (1984), ‘Applied general equilibrium models of taxation and international trade: an introduction and survey’, Journal of Economic Literature, 22, 1007– 1051. Sims, C. (1984), ‘Solving nonlinear stochastic equilibrium models backward’, University of Minnesota, manuscript. Smith, T. (1993), ‘Estimating nonlinear time series models using simulated VAR’, Journal of Applied Econometrics, 8, s63–s84. Tanner, M. and W.Wong (1987), ‘The calculation of posterior distributions by data augmentation’, Journal of the American Statistical Association, 87, 528–550. Tauchen, G. and R.Hussey (1991), ‘Quadrature based methods for obtaining approximate solutions to integral equations of nonlinear asset pricing models’, Econometrica, 59, 371– 397. Watson, M. (1993), ‘Measures of fit for calibrated models’, Journal of Political Economy, 101, 1011–1041. Weil, P. (1990), ‘The equity premium puzzle and the risk free puzzle’, Journal of Monetary Economics, 24, 401–421. Wolf, F. (1986), Meta-Analysis: Quantitative Methods for Research Synthesis, Beverly Hills, CA: Sage. © 1998 Selection and introduction © James E.Hartley, Kevin D.Hoover and Kevin D.Salyer; individual essays © their authors CHAPTER 19 355 INTERNATIONAL ECONOMIC REVIEW Vol. 36, No. 2, May 1995 SENSITIVITY ANALYSIS AND MODEL EVALUATION IN SIMULATED DYNAMIC GENERAL EQUILIBRIUM ECONOMIES* BY FABIO CANOVA1 This paper describes a Monte Carlo procedure to evaluate dynamic nonlinear general equilibrium macro models. The procedure makes the choice of parameters and the evaluation of the model less subjective than standard calibration techniques, it provides more general restrictions than estimation by simulation approaches and provides a way to conduct global sensitivity analysis for reasonable perturbations of the parameters. As an illustration the technique is applied to three examples involving different models and statistics. 1. INTRODUCTION A growing body of research in the applied macroeconomic literature uses simulation techniques to derive the time series properties of nonlinear stochastic general equilibrium models, to compare them to real world data and to evaluate policy options (see e.g. King, Plosser, and Rebelo 1988, or Cooley and Hansen 1990). In implementing numerical analyses of general equilibrium models, one has to overcome four hurdles. First, an economy must be specified and functional forms for its primitives selected. Second, a decision rule for the endogenous variables in terms of the exogenous (and predetermined) variables and of the parameters must be computed. Third, given the probability structure of the economy, values for the parameters must be chosen. Fourth, the closeness of functions of simulated and the actual data must be assessed in a metric which is relevant to the problem and policy conclusions, if any, should be drawn. While models are often specified with an eye to analytical tractability and there has been progress in developing techniques to numerically approximate unknown decision rules for the endogenous variables (see e.g. Sims 1984, Coleman 1989, Novales 1990, Baxter 1991, Tauchen and Hussey 1991, Judd 1992, Marcet 1992 and the January 1